The Inference Report

Token pruning is essential for enhancing the computational efficiency of vision-language models (VLMs), particularly for video-based tasks where temporal redundancy is prevalent. Prior approaches typically prune tokens either (1) within the vision transformer (ViT) exclusively for unimodal perception tasks such as action recognition and object segmentation, without adapting to downstream vision-language tasks; or (2) only within the LLM while leaving the ViT output intact, often requiring complex text-conditioned token selection mechanisms. In this paper, we introduce Spatio-Temporal Token Scoring (STTS), a simple and lightweight module that prunes vision tokens across both the ViT and the LLM without text conditioning or token merging, and is fully compatible with end-to-end training. By learning how to score temporally via an auxiliary loss and spatially via LLM downstream gradients, aided by our efficient packing algorithm, STTS prunes 50% of vision tokens throughout the entire architecture, resulting in a 62% improvement in efficiency during both training and inference with only a 0.7% drop in average performance across 13 short and long video QA tasks. Efficiency gains increase with more sampled frames per video. Applying test-time scaling for long-video QA further yields performance gains of 0.5-1% compared to the baseline. Overall, STTS represents a novel, simple yet effective technique for unified, architecture-wide vision token pruning.

Multimodal Large Language Models (MLLMs) have made impressive progress in connecting vision and language, but they still struggle with spatial understanding and viewpoint-aware reasoning. Recent efforts aim to augment the input representations with geometric cues rather than explicitly teaching models to reason in 3D space. We introduce Loc3R-VLM, a framework that equips 2D Vision-Language Models with advanced 3D understanding capabilities from monocular video input. Inspired by human spatial cognition, Loc3R-VLM relies on two joint objectives: global layout reconstruction to build a holistic representation of the scene structure, and explicit situation modeling to anchor egocentric perspective. These objectives provide direct spatial supervision that grounds both perception and language in a 3D context. To ensure geometric consistency and metric-scale alignment, we leverage lightweight camera pose priors extracted from a pre-trained 3D foundation model. Loc3R-VLM achieves state-of-the-art performance in language-based localization and outperforms existing 2D- and video-based approaches on situated and general 3D question-answering benchmarks, demonstrating that our spatial supervision framework enables strong 3D understanding. Project page: https://kevinqu7.github.io/loc3r-vlm

Building LLM-based agents has become increasingly important. Recent works on LLM-based agent self-evolution primarily record successful experiences as textual prompts or reflections, which cannot reliably guarantee efficient task re-execution in complex scenarios. We propose AgentFactory, a new self-evolution paradigm that preserves successful task solutions as executable subagent code rather than textual experience. Crucially, these subagents are continuously refined based on execution feedback, becoming increasingly robust and efficient as more tasks are encountered. Saved subagents are pure Python code with standardized documentation, enabling portability across any Python-capable system. We demonstrate that AgentFactory enables continuous capability accumulation: its library of executable subagents grows and improves over time, progressively reducing the effort required for similar tasks without manual intervention. Our implementation is open-sourced at https://github.com/zzatpku/AgentFactory, and our demonstration video is available at https://youtu.be/iKSsuAXJHW0.

Tokenization is a fundamental technique in the generative modeling of various modalities. In particular, it plays a critical role in autoregressive (AR) models, which have recently emerged as a compelling option for 3D generation. However, optimal tokenization of 3D shapes remains an open question. State-of-the-art (SOTA) methods primarily rely on geometric level-of-detail (LoD) hierarchies, originally designed for rendering and compression. These spatial hierarchies are often token-inefficient and lack semantic coherence for AR modeling. We propose Level-of-Semantics Tokenization (LoST), which orders tokens by semantic salience, such that early prefixes decode into complete, plausible shapes that possess principal semantics, while subsequent tokens refine instance-specific geometric and semantic details. To train LoST, we introduce Relational Inter-Distance Alignment (RIDA), a novel 3D semantic alignment loss that aligns the relational structure of the 3D shape latent space with that of the semantic DINO feature space. Experiments show that LoST achieves SOTA reconstruction, surpassing previous LoD-based 3D shape tokenizers by large margins on both geometric and semantic reconstruction metrics. Moreover, LoST achieves efficient, high-quality AR 3D generation and enables downstream tasks like semantic retrieval, while using only 0.1%-10% of the tokens needed by prior AR models.

Software vulnerabilities continue to grow in volume and remain difficult to detect in practice. Although learning-based vulnerability detection has progressed, existing benchmarks are largely function-centric and fail to capture realistic, executable, interprocedural settings. Recent repo-level security benchmarks demonstrate the importance of realistic environments, but their manual curation limits scale. This doctoral research proposes an automated benchmark generator that injects realistic vulnerabilities into real-world repositories and synthesizes reproducible proof-of-vulnerability (PoV) exploits, enabling precisely labeled datasets for training and evaluating repo-level vulnerability detection agents. We further investigate an adversarial co-evolution loop between injection and detection agents to improve robustness under realistic constraints.

AI coding agents can resolve real-world software issues, yet they frequently introduce regressions, breaking tests that previously passed. Current benchmarks focus almost exclusively on resolution rate, leaving regression behavior under-studied. This paper presents TDAD (Test-Driven Agentic Development), an open-source tool and benchmark methodology that combines abstract-syntax-tree (AST) based code-test graph construction with weighted impact analysis to surface the tests most likely affected by a proposed change. Evaluated on SWE-bench Verified with two local models (Qwen3-Coder 30B on 100 instances and Qwen3.5-35B-A3B on 25 instances), TDAD's GraphRAG workflow reduced test-level regressions by 70% (6.08% to 1.82%) and improved resolution from 24% to 32% when deployed as an agent skill. A surprising finding is that TDD prompting alone increased regressions (9.94%), revealing that smaller models benefit more from contextual information (which tests to verify) than from procedural instructions (how to do TDD). An autonomous auto-improvement loop raised resolution from 12% to 60% on a 10-instance subset with 0% regression. These findings suggest that for AI agent tool design, surfacing contextual information outperforms prescribing procedural workflows. All code, data, and logs are publicly available at https://github.com/pepealonso95/TDAD.

Orthogonalized-momentum optimizers such as Muon improve transformer training by approximately whitening/orthogonalizing matrix-valued momentum updates via a short polar-decomposition iteration. However, polar-factor approximations typically require multiple large matrix multiplications, and the resulting overhead can be substantial and hardware-dependent. We introduce MUD (MomentUm Decorrelation), a complementary whitening approach that replaces Muon's polar update with a triangular (Cholesky-like) whitening surrogate inspired by classical Gram--Schmidt and Gauss-Seidel ideas. We show that row-orthonormal matrices are fixed points of the MUD map, relate the inner step to symmetric Gauss-Seidel preconditioning of the Gram matrix, and prove quadratic local convergence near the fixed point. In terms of time-to-perplexity, MUD yields consistent 10-50\% wall-clock improvements over tuned AdamW and Muon in time-to-perplexity, typically converging slightly slower per step than Muon but with substantially lower optimizer overhead -- relative to Muon, MUD improves peak tokens/s by roughly $1.3-2.6\times$ across most settings and up to nearly $3\times$ on GPT-2 large on an A100. We also demonstrate training a ESM-2 150M protein language model, where MUD matches Muon-level validation perplexity in significantly less wall-clock time.

Robotics foundation models have demonstrated strong capabilities in executing natural language instructions across diverse tasks and environments. However, they remain largely data-driven and lack formal guarantees on safety and satisfaction of time-dependent specifications during deployment. In practice, robots often need to comply with operational constraints involving rich spatio-temporal requirements such as time-bounded goal visits, sequential objectives, and persistent safety conditions. In this work, we propose a specification-aware action distribution optimization framework that enforces a broad class of Signal Temporal Logic (STL) constraints during execution of a pretrained robotics foundation model without modifying its parameters. At each decision step, the method computes a minimally modified action distribution that satisfies a hard STL feasibility constraint by reasoning over the remaining horizon using forward dynamics propagation. We validate the proposed framework in simulation using a state-of-the-art robotics foundation model across multiple environments and complex specifications.

Handling gender across languages remains a persistent challenge for Machine Translation (MT) and Large Language Models (LLMs), especially when translating from gender-neutral languages into morphologically gendered ones, such as English to Italian. English largely omits grammatical gender, while Italian requires explicit agreement across multiple grammatical categories. This asymmetry often leads MT systems to default to masculine forms, reinforcing bias and reducing translation accuracy. To address this issue, we present the Contextual Gender Annotation (ConGA) framework, a linguistically grounded set of guidelines for word-level gender annotation. The scheme distinguishes between semantic gender in English through three tags, Masculine (M), Feminine (F), and Ambiguous (A), and grammatical gender realisation in Italian (Masculine (M), Feminine (F)), combined with entity-level identifiers for cross-sentence tracking. We apply ConGA to the gENder-IT dataset, creating a gold-standard resource for evaluating gender bias in translation. Our results reveal systematic masculine overuse and inconsistent feminine realisation, highlighting persistent limitations of current MT systems. By combining fine-grained linguistic annotation with quantitative evaluation, this work offers both a methodology and a benchmark for building more gender-aware and multilingual NLP systems.

While Large Language Models achieve state-of-the-art results across a wide range of NLP tasks, they remain prone to systematic biases. Among these, gender bias is particularly salient in MT, due to systematic differences across languages in whether and how gender is marked. As a result, translation often requires disambiguating implicit source signals into explicit gender-marked forms. In this context, standard benchmarks may capture broad disparities but fail to reflect the full complexity of gender bias in modern MT. In this paper, we extend recent frameworks on bias evaluation by: (i) introducing a novel measure coined "Prior Bias", capturing a model's default gender assumptions, and (ii) applying the framework to decoder-only MT models. Our results show that, despite their scale and state-of-the-art status, decoder-only models do not generally outperform encoder-decoder architectures on gender-specific metrics; however, post-training (e.g., instruction tuning) not only improves contextual awareness but also reduces the masculine Prior Bias.

Extending language models to video introduces two challenges: representation, where existing methods rely on lossy approximations, and long-context, where caption- or agent-based pipelines collapse video into text and lose visual fidelity. To overcome this, we introduce \textbf{VideoAtlas}, a task-agnostic environment to represent video as a hierarchical grid that is simultaneously lossless, navigable, scalable, caption- and preprocessing-free. An overview of the video is available at a glance, and any region can be recursively zoomed into, with the same visual representation used uniformly for the video, intermediate investigations, and the agent's memory, eliminating lossy text conversion end-to-end. This hierarchical structure ensures access depth grows only logarithmically with video length. For long-context, Recursive Language Models (RLMs) recently offered a powerful solution for long text, but extending them to visual domain requires a structured environment to recurse into, which \textbf{VideoAtlas} provides. \textbf{VideoAtlas} as a Markov Decision Process unlocks Video-RLM: a parallel Master-Worker architecture where a Master coordinates global exploration while Workers concurrently drill into assigned regions to accumulate lossless visual evidence. We demonstrate three key findings: (1)~logarithmic compute growth with video duration, further amplified by a 30-60\% multimodal cache hit rate arising from the grid's structural reuse. (2)~environment budgeting, where bounding the maximum exploration depth provides a principled compute-accuracy hyperparameter. (3)~emergent adaptive compute allocation that scales with question granularity. When scaling from 1-hour to 10-hour benchmarks, Video-RLM remains the most duration-robust method with minimal accuracy degradation, demonstrating that structured environment navigation is a viable and scalable paradigm for video understanding.

Rapid adaptation in complex control systems remains a central challenge in reinforcement learning. We introduce a framework in which policy and value functions share a low-dimensional coefficient vector - a goal embedding - that captures task identity and enables immediate adaptation to novel tasks without retraining representations. During pretraining, we jointly learn structured value bases and compatible policy bases through a bilinear actor-critic decomposition. The critic factorizes as Q = sum_k G_k(g) y_k(s,a), where G_k(g) is a goal-conditioned coefficient vector and y_k(s,a) are learned value basis functions. This multiplicative gating - where a context signal scales a set of state-dependent bases - is reminiscent of gain modulation observed in Layer 5 pyramidal neurons, where top-down inputs modulate the gain of sensory-driven responses without altering their tuning. Building on Successor Features, we extend the decomposition to the actor, which composes a set of primitive policies weighted by the same coefficients G_k(g). At test time the bases are frozen and G_k(g) is estimated zero-shot via a single forward pass, enabling immediate adaptation to novel tasks without any gradient update. We train a Soft Actor-Critic agent on the MuJoCo Ant environment under a multi-directional locomotion objective, requiring the agent to walk in eight directions specified as continuous goal vectors. The bilinear structure allows each policy head to specialize to a subset of directions, while the shared coefficient layer generalizes across them, accommodating novel directions by interpolating in goal embedding space. Our results suggest that shared low-dimensional goal embeddings offer a general mechanism for rapid, structured adaptation in high-dimensional control, and highlight a potentially biologically plausible principle for efficient transfer in complex reinforcement learning systems.

Converting pretrained attention modules such as grouped-query attention (GQA) into multi-head latent attention (MLA) can improve expressivity without increasing KV-cache cost, making it attractive for efficient inference. However, many practical conversion baselines rely on weight-only low-rank approximations (e.g., SVD-style initializations) and uniform rank allocation. They focus on minimizing the difference between weight matrices rather than on how those weights affect input activations, ignore the covariance structure of activations, and enforce uniform rank across layers, causing activation drift and degraded attention fidelity. To address these issues, we propose CARE, a Covariance-Aware, Rank-Enhanced MLA conversion pipeline under a fixed KV width. CARE introduces three key steps: (i) activation-preserving factorization, which aligns the approximation with the actual input activations rather than just the weights; (ii) adjusted-rank allocation, which spreads a fixed KV budget across layers by giving more capacity to layers that need it most; and (iii) KV-parity mapping, which reparameterizes the converted K and V to fit the MLA format while keeping the KV-cache size unchanged. Our method outperforms a uniform-rank SVD baseline on Qwen3-4B/30B-A3B-Instruct-2507 and Llama-3.1-8B/70B-Instruct, reducing one-shot perplexity by up to 215x and improving mean accuracy by up to 1.70x at matched KV budgets. With a brief post-SVD healing fine-tune, we fully recover the original model's accuracy.

In multilingual pretraining, the test loss of a pretrained model is heavily influenced by the proportion of each language in the pretraining data, namely the \textit{language mixture ratios}. Multilingual scaling laws can predict the test loss under different language mixture ratios and can therefore be used to estimate the optimal ratios. However, the current approaches to multilingual scaling laws do not measure the \textit{cross-lingual transfer} effect, resulting in suboptimal mixture ratios. In this paper, we consider multilingual pretraining as a cooperative game in which each language acts as a player that jointly contributes to pretraining, gaining the resulting reduction in test loss as the payoff. Consequently, from the perspective of cooperative game theory, we quantify the cross-lingual transfer from each language by its contribution in the game, and propose a game-theoretic multilingual scaling law called \textit{ShapleyLaw}. Our experiments show that ShapleyLaw outperforms baseline methods in model performance prediction and language mixture optimization.

Large language models (LLMs) exhibit latent multi-token prediction (MTP) capabilities despite being trained solely for next-token generation. We propose a simple, training-free MTP approach that probes an LLM using on-the-fly mask tokens drawn from its embedding space, enabling parallel prediction of future tokens without modifying model weights or relying on auxiliary draft models. Our method constructs a speculative token tree by sampling top-K candidates from mask-token logits and applies a lightweight pruning strategy to retain high-probability continuations. During decoding, candidate predictions are verified in parallel, resulting in lossless generation while substantially reducing the number of model calls and improving token throughput. Across benchmarks, our probing-based MTP consistently outperforms existing training-free baselines, increasing acceptance length by approximately 12\% on LLaMA3 and 8--12\% on Qwen3, and achieving throughput gains of up to 15--19\%. Finally, we provide theoretical insights and empirical evidence showing that decoder layers naturally align mask-token representations with next-token states, enabling accurate multi-step prediction without retraining or auxiliary models.

We consider a variant of sequential testing by betting where, at each time step, the statistician is presented with multiple data sources (arms) and obtains data by choosing one of the arms. We consider the composite global null hypothesis $\mathscr{P}$ that all arms are null in a certain sense (e.g. all dosages of a treatment are ineffective) and we are interested in rejecting $\mathscr{P}$ in favor of a composite alternative $\mathscr{Q}$ where at least one arm is non-null (e.g. there exists an effective treatment dosage). We posit an optimality desideratum that we describe informally as follows: even if several arms are non-null, we seek $e$-processes and sequential tests whose performance are as strong as the ones that have oracle knowledge about which arm generates the most evidence against $\mathscr{P}$. Formally, we generalize notions of log-optimality and expected rejection time optimality to more than one arm, obtaining matching lower and upper bounds for both. A key technical device in this optimality analysis is a modified upper-confidence-bound-like algorithm for unobservable but sufficiently "estimable" rewards. In the design of this algorithm, we derive nonasymptotic concentration inequalities for optimal wealth growth rates in the sense of Kelly [1956]. These may be of independent interest.

Accurate software energy measurement is critical for optimizing energy, yet existing profilers force a trade-off between measurement accuracy and overhead due to tight coupling with supported specific hardware or languages. We present CodeGreen, a modular energy measurement platform that decouples instrumentation from measurement via an asynchronous producer-consumer architecture. We implement a Native Energy Measurement Backend (NEMB) that polls hardware sensors (Intel RAPL, NVIDIA NVML, AMD ROCm) independently, while lightweight timestamp markers enable tunable granularity. CodeGreen leverages Tree-sitter AST queries for automated instrumentation across Python, C++, C, and Java, with straightforward extension to any Tree-sitter-supported grammar, enabling developers to target specific scopes (loops, methods, classes) without manual intervention. Validation against "Computer Language Benchmarks Game" demonstrates $R^2 = 0.9934$ correlation with RAPL ground truth and $R^2 = 0.9997$ energy-workload linearity. By bridging fine-grained measurement precision with cross-platform portability, CodeGreen enables practical algorithmic energy optimization across heterogeneous environments. Source code, video demonstration, and documentation for the tool are publicly available at: https://smart-dal.github.io/codegreen/.

Large language models (LLMs) contain billions of parameters, yet many exact values are not essential. We show that what matters most is the relative rank of weights-whether one connection is stronger or weaker than another-rather than precise magnitudes. To reduce the number of unique weight values, we apply weight clustering to pretrained models, replacing every weight matrix with K shared values from K-means. For Llama 3.1-8B-Instruct and SmolLM2-135M, reducing each matrix to only 16-64 distinct values preserves strong accuracy without retraining, providing a simple, training-free method to compress LLMs on disk. Optionally fine-tuning only the cluster means (centroids) recovers 30-40 percent of the remaining accuracy gap at minimal cost. We then systematically randomize cluster means while keeping assignments fixed. Scrambling the relative ranks of the clusters degrades quality sharply-perplexity can increase by orders of magnitude-even when global statistics such as mean and variance are preserved. In contrast, rank-preserving randomizations cause almost no loss at mid and late layers. On the other hand, when many layers are perturbed simultaneously, progressive layer-by-layer replacement reveals that scale drift-not rank distortion-is the dominant collapse mechanism; however, an affine correction w' = aw + b with a > 0 (which preserves both rank order and overall weight distribution) can substantially delay this drift. This rank-based perspective offers a new lens on model compression and robustness.

As large language models (LLMs) are deployed in multilingual settings, their safety behavior in culturally diverse, low-resource languages remains poorly understood. We present the first systematic evaluation of LLM safety across 12 Indic languages, spoken by over 1.2 billion people but underrepresented in LLM training data. Using a dataset of 6,000 culturally grounded prompts spanning caste, religion, gender, health, and politics, we assess 10 leading LLMs on translated variants of the prompt. Our analysis reveals significant safety drift: cross-language agreement is just 12.8\%, and \texttt{SAFE} rate variance exceeds 17\% across languages. Some models over-refuse benign prompts in low-resource scripts, overflag politically sensitive topics, while others fail to flag unsafe generations. We quantify these failures using prompt-level entropy, category bias scores, and multilingual consistency indices. Our findings highlight critical safety generalization gaps in multilingual LLMs and show that safety alignment does not transfer evenly across languages. We release \textsc{IndicSafe}, the first benchmark to enable culturally informed safety evaluation for Indic deployments, and advocate for language-aware alignment strategies grounded in regional harms.

Understanding the distance between human languages is central to linguistics, anthropology, and tracing human evolutionary history. Yet, while linguistics has long provided rich qualitative accounts of cross-linguistic variation, a unified and scalable quantitative approach to measuring language distance remains lacking. In this paper, we introduce a method that leverages pretrained multilingual language models as systematic instruments for linguistic measurement. Specifically, we show that the spontaneously emerged attention mechanisms of these models provide a robust, tokenization-agnostic measure of cross-linguistic distance, termed Attention Transport Distance (ATD). By treating attention matrices as probability distributions and measuring their geometric divergence via optimal transport, we quantify the representational distance between languages during translation. Applying ATD to a large and diverse set of languages, we demonstrate that the resulting distances recover established linguistic groupings with high fidelity and reveal patterns aligned with geographic and contact-induced relationships. Furthermore, incorporating ATD as a regularizer improves transfer performance in low-resource machine translation. Our results establish a principled foundation for testing linguistic hypotheses using artificial neural networks. This framework transforms multilingual models into powerful tools for quantitative linguistic discovery, facilitating more equitable multilingual AI.

Runtime verification has gained popularity as a lightweight approach for increasing assurance in systems under scrutiny. Performing runtime checks enables dynamic monitoring and alerts for unexpected behavior, thereby improving reliability and correctness. Actor-based systems present significant challenges for runtime verification. Properties frequently span multiple actors with complex causal dependencies, while nondeterministic message interleavings can obscure execution semantics. Moreover, most existing monitoring tools are designed for single-process behavior. This paper presents ACTORCHESTRA, a runtime verification framework for Erlang that automatically tracks causality across multi-actor interactions. The framework instruments Erlang systems that comply with OTP guidelines via targeted code injection. This method establishes the orchestration infrastructure required to track causal relationships between actors without requiring manual modifications to the target system. To ease the specification of multi-actor properties, the framework provides WALTZ, a specification language that automatically compiles properties into executable Erlang monitors that integrate with the instrumented system. Three case studies demonstrate ACTORCHESTRA's effectiveness in detecting complex behavioral violations in real-world actor systems. A performance evaluation quantifies the runtime overhead of the monitoring infrastructure and analyzes the trade-offs between added safety guarantees and execution costs.

Large language models (LLMs) and AI agents are increasingly integrated into enterprise systems to access internal databases and generate context-aware responses. While such integration improves productivity and decision support, the model outputs may inadvertently reveal sensitive information. Although many prior efforts focus on protecting the privacy of user prompts, relatively few studies consider privacy risks from the enterprise data perspective. Hence, this paper develops a probabilistic framework for analyzing privacy leakage in AI agents based on differential privacy. We model response generation as a stochastic mechanism that maps prompts and datasets to distributions over token sequences. Within this framework, we introduce token-level and message-level differential privacy and derive privacy bounds that relate privacy leakage to generation parameters such as temperature and message length. We further formulate a privacy-utility design problem that characterizes optimal temperature selection.

Understanding when learning is statistically possible yet computationally hard is a central challenge in high-dimensional statistics. In this work, we investigate this question in the context of single- and multi-index models, classes of functions widely studied as benchmarks to probe the ability of machine learning methods to discover features in high-dimensional data. Our main contribution is to show that a Noise Sensitivity Exponent (NSE) - a simple quantity determined by the activation function - governs the existence and magnitude of statistical-to-computational gaps within a broad regime of these models. We first establish that, in single-index models with large additive noise, the onset of a computational bottleneck is fully characterized by the NSE. We then demonstrate that the same exponent controls a statistical-computational gap in the specialization transition of large separable multi-index models, where individual components become learnable. Finally, in hierarchical multi-index models, we show that the NSE governs the optimal computational rate in which different directions are sequentially learned. Taken together, our results identify the NSE as a unifying property linking noise robustness, computational hardness, and feature specialization in high-dimensional learning.

Methodology bugs in scientific Python code produce plausible but incorrect results that traditional linters and static analysis tools cannot detect. Several research groups have built ML-specific linters, demonstrating that detection is feasible. Yet these tools share a sustainability problem: dependency on specific pylint or Python versions, limited packaging, and reliance on manual engineering for every new pattern. As AI-generated code increases the volume of scientific software, the need for automated methodology checking (such as detecting data leakage, incorrect cross-validation, and missing random seeds) grows. We present scicode-lint, whose two-tier architecture separates pattern design (frontier models at build time) from execution (small local model at runtime). Patterns are generated, not hand-coded; adapting to new library versions costs tokens, not engineering hours. On Kaggle notebooks with human-labeled ground truth, preprocessing leakage detection reaches 65% precision at 100% recall; on 38 published scientific papers applying AI/ML, precision is 62% (LLM-judged) with substantial variation across pattern categories; on a held-out paper set, precision is 54%. On controlled tests, scicode-lint achieves 97.7% accuracy across 66 patterns.

Post training quantization is essential for deploying large language models (LLMs) on resource constrained hardware, yet state of the art methods enforce uniform bit widths across layers, yielding suboptimal accuracy efficiency trade offs. We present RAMP (Reinforcement Adaptive Mixed Precision), an off policy Soft Actor Critic framework that learns per layer bit width assignments to minimize perplexity under a global bit budget. The policy conditions on an 11 dimensional embedding of activation statistics, weight properties, and structural descriptors, enabling zero shot transfer across model families and scales. To enable stable sub 4 bit quantization, we introduce Scale Folding, a preconditioning technique that migrates activation outliers into weights via per channel scaling and normalization layer compensation. A quality prioritized reward with asymmetric penalties and budget cliffs drives rapid convergence. On Llama 2 7B, RAMP achieves 5.54 perplexity at 3.68GB (3.65 effective bits), outperforming uniform 4 bit AWQ (5.60 at 3.90 GB) and GPTQ by 6% in size and 1% to3% in quality. Critically, a policy trained only on Llama 2 7B generalizes zero shot to Llama 2 13B and Mistral 7B, often surpassing target specific training, supporting the hypothesis that quantization sensitivity is primarily architectural. The HALO pipeline exports allocations to GGUF format for kernel free inference on CPUs, GPUs, and edge devices, retaining 99.5% of FP16 commonsense reasoning performance.

Helping people identify and pursue personally meaningful career goals at scale remains a key challenge in applied psychology. Career coaching can improve goal quality and attainment, but its cost and limited availability restrict access. Large language model (LLM)-based chatbots offer a scalable alternative, yet the psychological mechanisms by which they might support goal pursuit remain untested. Here we report a preregistered three-arm randomised controlled trial (N = 517) comparing an AI career coach ("Leon," powered by Claude Sonnet), a matched structured written questionnaire covering closely matched reflective topics, and a no-support control on goal progress at a two-week follow-up. The AI chatbot produced significantly higher goal progress than the control (d = 0.33, p = .016). Compared with the written-reflection condition, the AI did not significantly improve overall goal progress, but it increased perceived social accountability. In the preregistered mediation model, perceived accountability mediated the AI-over-questionnaire effect on goal progress (indirect effect = 0.15, 95% CI [0.04, 0.31]), whereas self-concordance did not. These findings suggest that AI-assisted goal setting can improve short-term goal progress, and that its clearest added value over structured self-reflection lies in increasing felt accountability.

Large language models (LLMs) are trained through multi-stage pipelines over heterogeneous data sources, yet developers lack a principled way to pinpoint the specific data responsible for an observed behavior. This lack of observability reduces debugging to reactive patching and makes failures prone to recur under distribution shift or subsequent model updates. To address this limitation, we propose DebugLM, a framework that equips LLMs with built-in data provenance, enabling them to explicitly trace the origins of their behaviors to specific training data sources. Specifically, the model learns to associate its responses with unique provenance tags that indicate the responsible dataset, empowering developers to precisely identify where undesirable behaviors are learned. Building on this capability, DebugLM further supports targeted test-time remediation, enabling developers to selectively trigger targeted refusal for specified data sources without retraining or modifying model parameters. Experiments demonstrate that DebugLM provides accurate behavior tracing in multi-stage training pipelines and effective test-time remediation while preserving the general utility of the model.

This work presents a multi-label classification framework for video capsule endoscopy (VCE) that addresses the extreme class imbalance inherent in the Galar dataset through a combination of architectural and optimization-level strategies. Our approach modifies BiomedCLIP, a biomedical vision-language foundation model, by replacing its standard multi-head self-attention with a differential attention mechanism that computes the difference between two softmax attention maps to suppress attention noise. To counteract the skewed label distribution, where pathological findings constitute less than 0.1% of all annotated frames, a sqrt-frequency weighted sampler, asymmetric focal loss, mixup regularization, and per-class threshold optimization are employed. Temporal coherence is enforced through median-filter smoothing and gap merging prior to event-level JSON generation. On the held-out RARE-VISION test set comprising three NaviCam examinations (161,025 frames), the pipeline achieves an overall temporal mAP@0.5 of 0.2456 and mAP@0.95 of 0.2353, with total inference completed in approximately 8.6 minutes on a single GPU.

Markov decision processes (MDPs) is viewed as an optimization of an objective function over certain linear operators over general function spaces. Using the well-established perturbation theory of linear operators, this viewpoint allows one to identify derivatives of the objective function as a function of the linear operators. This leads to generalization of many well-known results in reinforcement learning to cases with generate state and action spaces. Prior results of this type were only established in the finite-state finite-action MDP settings and in settings with certain linear function approximations. The framework also leads to new low-complexity PPO-type reinforcement learning algorithms for general state and action space MDPs.

Large Language Models (LLMs) have achieved unprecedented fluency but remain susceptible to "hallucinations" - the generation of factually incorrect or ungrounded content. This limitation is particularly critical in high-stakes domains where reliability is paramount. We propose a domain-grounded tiered retrieval and verification architecture designed to systematically intercept factual inaccuracies by shifting LLMs from stochastic pattern-matchers to verified truth-seekers. The proposed framework utilizes a four-phase, self-regulating pipeline implemented via LangGraph: (I) Intrinsic Verification with Early-Exit logic to optimize compute, (II) Adaptive Search Routing utilizing a Domain Detector to target subject-specific archives, (III) Corrective Document Grading (CRAG) to filter irrelevant context, and (IV) Extrinsic Regeneration followed by atomic claim-level verification. The system was evaluated across 650 queries from five diverse benchmarks: TimeQA v2, FreshQA v2, HaluEval General, MMLU Global Facts, and TruthfulQA. Empirical results demonstrate that the pipeline consistently outperforms zero-shot baselines across all environments. Win rates peaked at 83.7% in TimeQA v2 and 78.0% in MMLU Global Facts, confirming high efficacy in domains requiring granular temporal and numerical precision. Groundedness scores remained robustly stable between 78.8% and 86.4% across factual-answer rows. While the architecture provides a robust fail-safe for misinformation, a persistent failure mode of "False-Premise Overclaiming" was identified. These findings provide a detailed empirical characterization of multi-stage RAG behavior and suggest that future work should prioritize pre-retrieval "answerability" nodes to further bridge the reliability gap in conversational AI.

We propose RHYME-XT, an operator-learning framework for surrogate modeling of spatiotemporal control systems governed by input-affine nonlinear partial integro-differential equations (PIDEs) with localized rhythmic behavior. RHYME-XT uses a Galerkin projection to approximate the infinite-dimensional PIDE on a learned finite-dimensional subspace with spatial basis functions parameterized by a neural network. This yields a projected system of ODEs driven by projected inputs. Instead of integrating this non-autonomous system, we directly learn its flow map using an architecture for learning flow functions, avoiding costly computations while obtaining a continuous-time and discretization-invariant representation. Experiments on a neural field PIDE show that RHYME-XT outperforms a state-of-the-art neural operator and is able to transfer knowledge effectively across models trained on different datasets, through a fine-tuning process.

Automating the development of machine learning algorithms has the potential to unlock new breakthroughs. However, our ability to improve and evaluate algorithm discovery systems has thus far been limited by existing task suites. They suffer from many issues, such as: poor evaluation methodologies; data contamination; and containing saturated or very similar problems. Here, we introduce DiscoGen, a procedural generator of algorithm discovery tasks for machine learning, such as developing optimisers for reinforcement learning or loss functions for image classification. Motivated by the success of procedural generation in reinforcement learning, DiscoGen spans millions of tasks of varying difficulty and complexity from a range of machine learning fields. These tasks are specified by a small number of configuration parameters and can be used to optimise algorithm discovery agents (ADAs). We present DiscoBench, a benchmark consisting of a fixed, small subset of DiscoGen tasks for principled evaluation of ADAs. Finally, we propose a number of ambitious, impactful research directions enabled by DiscoGen, in addition to experiments demonstrating its use for prompt optimisation of an ADA. DiscoGen is released open-source at https://github.com/AlexGoldie/discogen.

Modern seismic and volcanic monitoring is increasingly shaped by continuous, multi-sensor observations and by the need to extract actionable information from nonstationary, noisy wavefields. In this context, machine learning has moved from a research curiosity to a practical ingredient of processing chains for detection, phase picking, classification, denoising, and anomaly tracking. However, improved accuracy on a fixed dataset is not sufficient for operational use. Models must remain reliable under domain shift (new stations, changing noise, evolving volcanic activity), provide uncertainty that supports decision-making, and connect their outputs to physically meaningful constraints. This paper surveys and organizes recent ML approaches for seismic and volcanic signal analysis, highlighting where classical signal processing provides indispensable inductive bias, how self-supervision and generative modeling can reduce dependence on labels, and which evaluation protocols best reflect transfer across regions. We conclude with open challenges for robust, interpretable, and maintainable AI-assisted monitoring.

Verbal confidence -- prompting LLMs to state their confidence as a number or category -- is widely used to extract uncertainty estimates from black-box models. However, how LLMs internally generate such scores remains unknown. We address two questions: first, when confidence is computed - just-in-time when requested, or automatically during answer generation and cached for later retrieval; and second, what verbal confidence represents - token log-probabilities, or a richer evaluation of answer quality? Focusing on Gemma 3 27B and Qwen 2.5 7B, we provide convergent evidence for cached retrieval. Activation steering, patching, noising, and swap experiments reveal that confidence representations emerge at answer-adjacent positions before appearing at the verbalization site. Attention blocking pinpoints the information flow: confidence is gathered from answer tokens, cached at the first post-answer position, then retrieved for output. Critically, linear probing and variance partitioning reveal that these cached representations explain substantial variance in verbal confidence beyond token log-probabilities, suggesting a richer answer-quality evaluation rather than a simple fluency readout. These findings demonstrate that verbal confidence reflects automatic, sophisticated self-evaluation -- not post-hoc reconstruction -- with implications for understanding metacognition in LLMs and improving calibration.

Existing human value datasets do not directly support value understanding in factual news: many are actor-agnostic, rely on isolated utterances or synthetic scenarios, and lack explicit event structure or value direction. We present \textbf{NEVU} (\textbf{N}ews \textbf{E}vent-centric \textbf{V}alue \textbf{U}nderstanding), a benchmark for \emph{actor-conditioned}, \emph{event-centric}, and \emph{direction-aware} human value recognition in factual news. NEVU evaluates whether models can identify value cues, attribute them to the correct actor, and determine value direction from grounded evidence. Built from 2{,}865 English news articles, NEVU organizes annotations at four semantic unit levels (\textbf{Subevent}, \textbf{behavior-based composite event}, \textbf{story-based composite event}, and \textbf{Article}) and labels \mbox{(unit, actor)} pairs for fine-grained evaluation across local and composite contexts. The annotations are produced through an LLM-assisted pipeline with staged verification and targeted human auditing. Using a hierarchical value space with \textbf{54} fine-grained values and \textbf{20} coarse-grained categories, NEVU covers 45{,}793 unit--actor pairs and 168{,}061 directed value instances. We provide unified baselines for proprietary and open-source LLMs, and find that lightweight adaptation (LoRA) consistently improves open-source models, showing that although NEVU is designed primarily as a benchmark, it also supports supervised adaptation beyond prompting-only evaluation. Data availability is described in Appendix~\ref{app:data_code_availability}.

During conversational interactions, humans subconsciously engage in concurrent thinking while listening to a speaker. Although this internal cognitive processing may not always manifest as explicit linguistic structures, it is instrumental in formulating high-quality responses. Inspired by this cognitive phenomenon, we propose a novel Full-duplex LAtent and Internal Reasoning method named FLAIR that conducts latent thinking simultaneously with speech perception. Unlike conventional "thinking" mechanisms in NLP, which require post-hoc generation, our approach aligns seamlessly with spoken dialogue systems: during the user's speaking phase, it recursively feeds the latent embedding output from the previous step into the next step, enabling continuous reasoning that strictly adheres to causality without introducing additional latency. To enable this latent reasoning, we design an Evidence Lower Bound-based objective that supports efficient supervised finetuning via teacher forcing, circumventing the need for explicit reasoning annotations. Experiments demonstrate the effectiveness of this think-while-listening design, which achieves competitive results on a range of speech benchmarks. Furthermore, FLAIR robustly handles conversational dynamics and attains competitive performance on full-duplex interaction metrics.

Physics-informed machine learning surrogates are increasingly explored to accelerate dynamic simulation of generators, converters, and other power grid components. The key question, however, is not only whether a surrogate matches a stand-alone component model on average, but whether it remains accurate after insertion into a differential-algebraic simulator, where the surrogate outputs enter the algebraic equations coupling the component to the rest of the system. This paper formulates that in-simulator use as a verification and validation (V\&V) problem. A finite-horizon bound is derived that links allowable component-output error to algebraic-coupling sensitivity, dynamic error amplification, and the simulation horizon. Two complementary settings are then studied: model-based verification against a reference component solver, and data-based validation through conformal calibration of the component-output variables exchanged with the simulator. The framework is general, but the case study focuses on physics-informed neural-network surrogates of second-, fourth-, and sixth-order synchronous-machine models. Results show that good stand-alone surrogate accuracy does not by itself guarantee accurate in-simulator behavior, that the largest discrepancies concentrate in stressed operating regions, and that small equation residuals do not necessarily imply small state-trajectory errors.

Diffusion models and flow matching have become a cornerstone of robotic imitation learning, yet they suffer from a structural inefficiency where inference is often bound to a fixed integration schedule that is agnostic to state complexity. This paradigm forces the policy to expend the same computational budget on trivial motions as it does on complex tasks. We introduce Generative Control as Optimization (GeCO), a time-unconditional framework that transforms action synthesis from trajectory integration into iterative optimization. GeCO learns a stationary velocity field in the action-sequence space where expert behaviors form stable attractors. Consequently, test-time inference becomes an adaptive process that allocates computation based on convergence--exiting early for simple states while refining longer for difficult ones. Furthermore, this stationary geometry yields an intrinsic, training-free safety signal, as the field norm at the optimized action serves as a robust out-of-distribution (OOD) detector, remaining low for in-distribution states while significantly increasing for anomalies. We validate GeCO on standard simulation benchmarks and demonstrate seamless scaling to pi0-series Vision-Language-Action (VLA) models. As a plug-and-play replacement for standard flow-matching heads, GeCO improves success rates and efficiency with an optimization-native mechanism for safe deployment. Video and code can be found at https://hrh6666.github.io/GeCO/

Benchmarks for large language models (LLMs) have progressed from snippet-level function generation to repository-level issue resolution, yet they overwhelmingly target implementation correctness. Software architecture tasks remain under-specified and difficult to compare across models, despite their central role in maintaining and evolving complex systems. We present ArchBench, the first unified platform for benchmarking LLM capabilities on software architecture tasks. ArchBench provides a command-line tool with a standardized pipeline for dataset download, inference with trajectory logging, and automated evaluation, alongside a public web interface with an interactive leaderboard. The platform is built around a plugin architecture where each task is a self-contained module, making it straightforward for the community to contribute new architectural tasks and evaluation results. We use the term LLMs broadly to encompass generative AI (GenAI) solutions for software engineering, including both standalone models and LLM-based coding agents equipped with tools. Both the CLI tool and the web platform are openly available to support reproducible research and community-driven growth of architectural benchmarking.

In this work, we probe the ability of a language model to demonstrate spatial reasoning from unstructured text, mimicking human capabilities and automating a process that benefits many downstream media applications. Concretely, we study the narrative-to-play task: inferring stage-play layouts (scenes, speaker positions, movements, and room types) from text that lacks explicit spatial, positional, or relational cues. We then introduce a dramaturgy-inspired deterministic evaluation suite and, finally, a training and inference recipe that combines rejection SFT using Best-of-N sampling with RL from verifiable rewards via GRPO. Experiments on a text-only corpus of classical English literature demonstrate improvements over vanilla models across multiple metrics (character attribution, spatial plausibility, and movement economy), as well as alignment with an LLM-as-a-judge and subjective human preferences.

LLM agents often fail in closed-world embodied environments because actions must satisfy strict preconditions -- such as location, inventory, and container states -- and failure feedback is sparse. We identify two structurally coupled failure modes: (P1) invalid action generation and (P2) state drift, each amplifying the other in a degenerative cycle. We present RPMS, a conflict-managed architecture that enforces action feasibility via structured rule retrieval, gates memory applicability via a lightweight belief state, and resolves conflicts between the two sources via rules-first arbitration. On ALFWorld (134 unseen tasks), RPMS achieves 59.7% single-trial success with Llama 3.1 8B (+23.9 pp over baseline) and 98.5% with Claude Sonnet 4.5 (+11.9 pp); of the 8B gain, rule retrieval alone contributes +14.9 pp (statistically significant), making it the dominant factor. A key finding is that episodic memory is conditionally useful: it harms performance on some task types when used without grounding, but becomes a stable net positive once filtered by current state and constrained by explicit action rules. Adapting RPMS to ScienceWorld with GPT-4 yields consistent gains across all ablation conditions (avg. score 54.0 vs. 44.9 for the ReAct baseline), providing transfer evidence that the core mechanisms hold across structurally distinct environments.

A prerequisite for coding agents to perform tasks on large repositories is code localization - the identification of relevant files, classes, and functions to work on. While repository-level code localization has been performed using embedding-based retrieval approaches such as vector search, recent work has focused on developing agents to localize relevant code either as a standalone precursor to or interleaved with performing actual work. Most prior methods on agentic code search equip the agent with complex, specialized tools, such as repository graphs derived from static analysis. In this paper, we demonstrate that, with an effective reinforcement learning recipe, a coding agent equipped with nothing more than a standard Unix terminal can be trained to achieve strong results. Our experiments on three benchmarks (SWE-Bench Verified, Pro, and Lite) reveal that our models consistently achieve superior or competitive performance over 2-18x larger base and post-trained LLMs and sometimes approach performance provided by closed models like Claude Sonnet, even when using specialized scaffolds. Our work particularly focuses on techniques for re-purposing existing coding agent environments for code search, reward design, and RL optimization. We release the resulting model family, CodeScout, along with all our code and data for the community to build upon.

Multimodal Automated Program Repair (MAPR) extends traditional program repair by requiring models to jointly reason over source code, textual issue descriptions, and visual artifacts such as GUI screenshots. While recent LLM-based repair systems have shown promising results, existing approaches face several limitations: rigid workflow pipelines restrict exploration during debugging, visual reasoning is often performed over full-page screenshots without localized grounding, and failed repair attempts are rarely transformed into reusable knowledge. To address these challenges, we propose FailureMem, a multimodal repair framework that integrates three key mechanisms: a hybrid workflow-agent architecture that balances structured localization with flexible reasoning, active perception tools that enable region-level visual grounding, and a Failure Memory Bank that converts past repair attempts into reusable guidance. Experiments on SWE-bench Multimodal demonstrate FailureMem improves the resolved rate over GUIRepair by 3.7%.

Tensegrity structures possess intrinsic geometric symmetries that govern their dynamic behavior. However, most existing physics-informed neural network (PINN) approaches for tensegrity dynamics do not explicitly exploit these symmetries, leading to high computational complexity and unstable optimization. In this work, we propose a symmetry-reduced physics-informed neural network (SymPINN) framework that embeds group-theory-based symmetry directly into both the solution expression and the neural network architecture to predict tensegrity dynamics. By decomposing nodes into symmetry orbits and representing free nodal coordinates using a symmetry basis, the proposed method constructs a reduced coordinate representation that preserves geometric symmetry of the structure. The full coordinates are then recovered via symmetry transformations of the reduced solution learned by the network, ensuring that the predicted configurations automatically satisfy the symmetry constraints. In this framework, equivariance is enforced through orbit-based coordinate generation, symmetry-consistent message passing, and physics residual constraints. In addition, SymPINN improves training effectiveness by encoding initial conditions as hard constraints, incorporating Fourier feature encoding to enhance the representation of dynamic motions, and employing a two-stage optimization strategy. Extensive numerical experiments on symmetric T-bars and lander structures demonstrate significantly improved prediction accuracy and computational efficiency compared to standard physics-informed models, indicating the great potential of symmetry-aware learning for structure-preserving modeling of tensegrity dynamics.

Understanding the internal functional organization of Large Language Models (LLMs) is crucial for improving their trustworthiness and performance. However, how LLMs organize different functions into modules remains highly unexplored. To bridge this gap, we formulate a functional module discovery problem and propose an Unsupervised LLM Cross-layer MOdule Discovery (ULCMOD) framework that simultaneously disentangles the large set of neurons in the entire LLM into modules while discovering the topics of input samples related to these modules. Our framework introduces a novel objective function and an efficient Iterative Decoupling (IterD) algorithm. Extensive experiments show that our method discovers high-quality, disentangled modules that capture more meaningful semantic information and achieve superior performance in various downstream tasks. Moreover, our qualitative analysis reveals that the discovered modules show semantic coherence, correspond to interpretable specializations, and a clear spatial and hierarchical organization within the LLM. Our work provides a novel tool for interpreting the functional modules of LLMs, filling a critical blank in LLM's interpretability research.

Large audio language models (LALMs) can answer questions about speech, music, and environmental sounds, yet their internal reasoning is largely opaque and difficult to validate. We describe TalTech's solution to the Agent Track of the Interspeech 2026 Audio Reasoning Challenge, in which systems are evaluated on reasoning process quality, specifically the factual accuracy, logical soundness, and completeness of their reasoning chains. Our multi-source ensemble pipeline uses two LALMs that generate independent observations, while a separate text-only reasoning model cross-checks these against outputs from 25 acoustic tools organized into reliability tiers. By grounding every inference step in explicit, reliability-tagged evidence, the system produces dense, verifiable reasoning chains. Our system ranked first in the challenge, outperforming all competing systems by a wide margin in challenge's reasoning quality metric.

Contextual embeddings generated by LLMs exhibit strong positional inductive biases, which can limit their ability to fully capture long-range, order-sensitive dependencies in highly structured source code. Consequently, how to further refine and enhance LLM embeddings for improved code understanding remains an open research question. To address this gap, we propose a hybrid LLM-RNN framework that reinforces LLM-generated contextual embeddings with a sequential RNN architecture. The embeddings reprocessing step aims to reinforce sequential semantics and strengthen order-aware dependencies inherent in source code. We evaluate the proposed hybrid models on both benchmark and real-world coding datasets. The experimental results show that the RoBERTa-BiGRU and CodeBERT-GRU models achieved accuracies of 66.40% and 66.03%, respectively, on the defect detection benchmark dataset, representing improvements of approximately 5.35% and 3.95% over the standalone RoBERTa and CodeBERT models. Furthermore, the CodeT5-GRU and CodeT5+-BiGRU models achieved accuracies of 67.90% and 67.79%, respectively, surpassing their base models and outperforming RoBERTa-BiGRU and CodeBERT-GRU by a notable margin. In addition, CodeT5-GRU model attains weighted and macro F1-scores of 67.18% and 67.00%, respectively, on the same dataset. Extensive experiments across three real-world datasets further demonstrate consistent and statistically significant improvements over standalone LLMs. Overall, our findings indicate that reprocessing contextual embeddings with RNN architectures enhances code understanding performance in LLM-based models.

Federated reinforcement learning typically aggregates value functions or policies by parameter averaging, which emphasizes expected return and can obscure statistical multimodality and tail behavior that matter in safety-critical settings. We formalize federated distributional reinforcement learning (FedDistRL), where clients parametrize quantile value function critics and federate these networks only. We also propose TR-FedDistRL, which builds a per client, risk-aware Wasserstein barycenter over a temporal buffer. This local barycenter provides a reference region to constrain the parameter averaged critic, ensuring necessary distributional information is not averaged out during the federation process. The distributional trust region is implemented as a shrink-squash step around this reference. Under fixed-policy evaluation, the feasibility map is nonexpansive and the update is contractive in a probe-set Wasserstein metric under evaluation. Experiments on a bandit, multi-agent gridworld, and continuous highway environment show reduced mean-smearing, improved safety proxies (catastrophe/accident rate), and lower critic/policy drift versus mean-oriented and non-federated baselines.

Multi-step reasoning improves the capabilities of large language models (LLMs) but increases the risk of errors propagating through intermediate steps. Process reward models (PRMs) mitigate this by scoring each step individually, enabling fine-grained supervision and improved reliability. Existing methods for training PRMs rely on costly human annotations or computationally intensive automatic labeling. We propose a novel approach to automatically generate step-level labels using Information Theory. Our method estimates how each reasoning step affects the likelihood of the correct answer, providing a signal of step quality. Importantly, it reduces computational complexity to $\mathcal{O}(N)$, improving over the previous $\mathcal{O}(N \log N)$ methods. We demonstrate that these labels enable effective chain-of-thought selection in best-of-$K$ evaluation settings across diverse reasoning benchmarks, including mathematics, Python programming, SQL, and scientific question answering. This work enables scalable and efficient supervision of LLM reasoning, particularly for tasks where error propagation is critical.

Recent video diffusion models achieve high-quality generation through recurrent frame processing where each frame generation depends on previous frames. However, this recurrent mechanism means that training such models in the pixel domain incurs prohibitive memory costs, as activations accumulate across the entire video sequence. This fundamental limitation also makes fine-tuning these models with pixel-wise losses computationally intractable for long or high-resolution videos. This paper introduces ChopGrad, a truncated backpropagation scheme for video decoding, limiting gradient computation to local frame windows while maintaining global consistency. We provide a theoretical analysis of this approximation and show that it enables efficient fine-tuning with frame-wise losses. ChopGrad reduces training memory from scaling linearly with the number of video frames (full backpropagation) to constant memory, and compares favorably to existing state-of-the-art video diffusion models across a suite of conditional video generation tasks with pixel-wise losses, including video super-resolution, video inpainting, video enhancement of neural-rendered scenes, and controlled driving video generation.

Transformer-based language models are widely deployed for reasoning, yet their behavior under inference-time stochasticity remains underexplored. While dropout is common during training, its inference-time effects via Monte Carlo sampling lack systematic evaluation across architectures, limiting understanding of model reliability in uncertainty-aware applications. This work analyzes dropout-induced variability across 19 transformer models using MC Dropout with 100 stochastic forward passes per sample. Dropout robustness is defined as maintaining high accuracy and stable predictions under stochastic inference, measured by standard deviation of per-run accuracies. A cognitive decomposition framework disentangles performance into memory and reasoning components. Experiments span five dropout configurations yielding 95 unique evaluations on 1,000 samples. Results reveal substantial architectural variation. Smaller models demonstrate perfect prediction stability while medium-sized models exhibit notable volatility. Mid-sized models achieve the best overall performance; larger models excel at memory tasks. Critically, 53% of models suffer severe accuracy degradation under baseline MC Dropout, with task-specialized models losing up to 24 percentage points, indicating unsuitability for uncertainty quantification in these architectures. Asymmetric effects emerge: high dropout reduces memory accuracy by 27 percentage points while reasoning degrades only 1 point, suggesting memory tasks rely on stable representations that dropout disrupts. 84% of models demonstrate memory-biased performance. This provides the first comprehensive MC Dropout benchmark for transformers, revealing dropout robustness is architecture-dependent and uncorrelated with scale. The cognitive profiling framework offers actionable guidance for model selection in uncertainty-aware applications.

Large Vision Language Models (LVLMs) have achieved remarkable success in a range of downstream tasks that require multimodal interaction, but their capabilities come with substantial computational and memory overhead, which hinders practical deployment. Among numerous acceleration techniques, post-training quantization is a popular and effective strategy for reducing memory cost and accelerating inference. However, existing LVLM quantization methods typically measure token sensitivity at the modality level, which fails to capture the complex cross-token interactions and falls short in quantitatively measuring the quantization error at the token level. As tokens interact within the model, the distinction between modalities gradually diminishes, suggesting the need for fine-grained calibration. Inspired by axiomatic attribution in mechanistic interpretability, we introduce a fine-grained quantization strategy on Quantization-aware Integrated Gradients (QIG), which leverages integrated gradients to quantitatively evaluate token sensitivity and push the granularity from modality level to token level, reflecting both inter-modality and intra-modality dynamics. Extensive experiments on multiple LVLMs under both W4A8 and W3A16 settings show that our method improves accuracy across models and benchmarks with negligible latency overhead. For example, under 3-bit weight-only quantization, our method improves the average accuracy of LLaVA-onevision-7B by 1.60%, reducing the gap to its full-precision counterpart to only 1.33%. The code is available at https://github.com/ucas-xiang/QIG.

Video generative models are increasingly used as world models for robotics, where a model generates a future visual rollout conditioned on the current observation and task instruction, and an inverse dynamics model (IDM) converts the generated frames into executable robot actions. However, current video world models lack explicit executability constraints. As a result, visually coherent rollouts may still violate rigid-body and kinematic consistency, producing unstable or infeasible control commands when decoded by an IDM. We refer to this mismatch between visual generation and physically executable control as the executability gap. While this gap can be mitigated at inference time using techniques such as rejection sampling, such approaches are inefficient due to the high cost of video generation. In this paper, we leverage the executability gap as a training signal and introduce Executable Video Alignment (EVA), a reinforcement-learning post-training framework for aligning video world models. EVA trains an inverse dynamics model on real robot trajectories and repurposes it as a reward model that evaluates generated videos through the action sequences they induce, encouraging smooth motions measured by velocity, acceleration, and jerk while penalizing actions that violate embodiment constraints. Importantly, the reward remains informative even when generated videos contain severe visual artifacts, since such artifacts typically translate into unstable or out-of-bound actions. Experiments on the RoboTwin benchmark and a real bimanual robot show that EVA reduces embodiment-specific artifacts in generated rollouts and improves downstream task execution success.

In practice, machine learning methods commonly require anomaly detection (AD) to filter inputs or detect distributional shifts. Typically, this is implemented by running a separate AD model alongside the primary model. However, this separation ignores the fact that the primary model already encodes substantial information about the target distribution. In this paper, we introduce On-Model AD, a setting for anomaly detection that explicitly leverages access to a related machine learning model. Within this setting, we propose RangeAD, an algorithm that utilizes neuron-wise output ranges derived from the primary model. RangeAD achieves superior performance even on high-dimensional tasks while incurring substantially lower inference costs. Our results demonstrate the potential of the On-Model AD setting as a practical framework for efficient anomaly detection.

Integrating quantum mechanics into drug discovery marks a decisive shift from empirical trial-and-error toward quantitative precision. However, the prohibitive cost of ab initio molecular dynamics has historically forced a compromise between chemical accuracy and computational scalability. This paper identifies the convergence of High-Performance Computing (HPC), Machine Learning (ML), and Quantum Computing (QC) as the definitive solution to this bottleneck. While ML foundation models, such as FeNNix-Bio1, enable quantum-accurate simulations, they remain tethered to the inherent limits of classical data generation. We detail how High-Performance Quantum Computing (HPQC), utilizing hybrid QPU-GPU architectures, will serve as the ultimate accelerator for quantum chemistry data. By leveraging Hilbert space mapping, these systems can achieve true chemical accuracy while bypassing the heuristics of classical approximations. We show how this tripartite convergence optimizes the drug discovery pipeline, spanning from initial system preparation to ML-driven, high-fidelity simulations. Finally, we position quantum-enhanced sampling as the beyond GPU frontier for modeling reactive cellular systems and pioneering next-generation materials.

Enterprise AI deploys dozens of autonomous agent nodes across workflows, each acting on the same entities with no shared memory and no common governance. We identify five structural challenges arising from this memory governance gap: memory silos across agent workflows; governance fragmentation across teams and tools; unstructured memories unusable by downstream systems; redundant context delivery in autonomous multi-step executions; and silent quality degradation without feedback loops. We present Governed Memory, a shared memory and governance layer addressing this gap through four mechanisms: a dual memory model combining open-set atomic facts with schema-enforced typed properties; tiered governance routing with progressive context delivery; reflection-bounded retrieval with entity-scoped isolation; and a closed-loop schema lifecycle with AI-assisted authoring and automated per-property refinement. We validate each mechanism through controlled experiments (N=250, five content types): 99.6% fact recall with complementary dual-modality coverage; 92% governance routing precision; 50% token reduction from progressive delivery; zero cross-entity leakage across 500 adversarial queries; 100% adversarial governance compliance; and output quality saturation at approximately seven governed memories per entity. On the LoCoMo benchmark, the architecture achieves 74.8% overall accuracy, confirming that governance and schema enforcement impose no retrieval quality penalty. The system is in production at Personize.ai.

In this paper, we study total variation (TV)-regularized training of infinite-width shallow ReLU neural networks, formulated as a convex optimization problem over measures on the unit sphere. Our approach leverages the duality theory of TV-regularized optimization problems to establish rigorous guarantees on the sparsity of the solutions to the training problem. Our analysis further characterizes how and when this sparsity persists in a low noise regime and for small regularization parameter. The key observation that motivates our analysis is that, for ReLU activations, the associated dual certificate is piecewise linear in the weight space. Its linearity regions, which we name dual regions, are determined by the activation patterns of the data via the induced hyperplane arrangement. Taking advantage of this structure, we prove that, on each dual region, the dual certificate admits at most one extreme value. As a consequence, the support of any minimizer is finite, and its cardinality can be bounded from above by a constant depending only on the geometry of the data-induced hyperplane arrangement. Then, we further investigate sufficient conditions ensuring uniqueness of such sparse solution. Finally, under a suitable non-degeneracy condition on the dual certificate along the boundaries of the dual regions, we prove that in the presence of low label noise and for small regularization parameter, solutions to the training problem remain sparse with the same number of Dirac deltas. Additionally, their location and the amplitudes converge, and, in case the locations lie in the interior of a dual region, the convergence happens with a rate that depends linearly on the noise and the regularization parameter.

We developed a multi-label gastrointestinal video analysis pipeline based on a ResNet-50 frame classifier followed by anatomy-guided temporal event decoding. The system predicts 17 labels, including 5 anatomy classes and 12 pathology classes, from frames resized to 336x336. A major challenge was severe class imbalance, particularly for rare pathology labels. To address this, we used clipped class-wise positive weighting in the training loss, which improved rare-class learning while maintaining stable optimization. At the temporal stage, we found that direct frame-to-event conversion produced fragmented mismatches with the official ground truth. The final submission therefore combined GT-style framewise event composition, anatomy vote smoothing, and anatomy-based pathology gating with a conservative hysteresis decoder. This design improved the final temporal mAP from 0.3801 to 0.4303 on the challenge test set.

Large language models increasingly serve as persistent knowledge workers, with in-context memory - facts stored in the prompt - as the default strategy. We benchmark in-context memory against Knowledge Objects (KOs), discrete hash-addressed tuples with O(1) retrieval. Within the context window, Claude Sonnet 4.5 achieves 100% exact-match accuracy from 10 to 7,000 facts (97.5% of its 200K window). However, production deployment reveals three failure modes: capacity limits (prompts overflow at 8,000 facts), compaction loss (summarization destroys 60% of facts), and goal drift (cascading compaction erodes 54% of project constraints while the model continues with full confidence). KOs achieve 100% accuracy across all conditions at 252x lower cost. On multi-hop reasoning, KOs reach 78.9% versus 31.6% for in-context. Cross-model replication across four frontier models confirms compaction loss is architectural, not model-specific. We additionally show that embedding retrieval fails on adversarial facts (20% precision at 1) and that neural memory (Titans) stores facts but fails to retrieve them on demand. We introduce density-adaptive retrieval as a switching mechanism and release the benchmark suite.

Label-free reinforcement learning enables large language models to improve reasoning capabilities without ground-truth supervision, typically by treating majority-voted answers as pseudo-labels. However, we identify a critical failure mode: as training maximizes self-consistency, output diversity collapses, causing the model to confidently reinforce systematic errors that evade detection. We term this the consensus trap. To escape it, we propose CoVerRL, a framework where a single model alternates between generator and verifier roles, with each capability bootstrapping the other. Majority voting provides noisy but informative supervision for training the verifier, while the improving verifier progressively filters self-consistent errors from pseudo-labels. This co-evolution creates a virtuous cycle that maintains high reward accuracy throughout training. Experiments across Qwen and Llama model families demonstrate that CoVerRL outperforms label-free baselines by 4.7-5.9\% on mathematical reasoning benchmarks. Moreover, self-verification accuracy improves from around 55\% to over 85\%, confirming that both capabilities genuinely co-evolve.

Attention sinks and massive activations are recurring and closely related phenomena in Transformer models. Existing studies have largely focused on the forward pass, making it unclear whether their connection is direct or mediated by a training-time mechanism. We study this question from the perspective of backpropagation. Empirically and theoretically, we show that under causal mask, attention sinks can induce pronounced gradient concentration, which we term gradient sinks. Furthermore, in pre-norm architectures with RMSNorm, massive activations can be understood as an adaptive response to this localized gradient pressure during training. To test this hypothesis, we introduce V-scale, a modification that adjusts value-path backpropagated gradients. In pretrained V-scale models, attention sinks are preserved whereas massive activations are suppressed. These results support the interpretation that gradient sink is a key training-time mediator linking attention sinks and massive activations.

We propose to model parallel streams of data, such as overlapped speech, using shuffles. Specifically, this paper shows how the shuffle product and partial order finite-state automata (FSAs) can be used for alignment and speaker-attributed transcription of overlapped speech. We train using the total score on these FSAs as a loss function, marginalizing over all possible serializations of overlapping sequences at subword, word, and phrase levels. To reduce graph size, we impose temporal constraints by constructing partial order FSAs. We address speaker attribution by modeling (token, speaker) tuples directly. Viterbi alignment through the shuffle product FSA directly enables one-pass alignment. We evaluate performance on synthetic LibriSpeech overlaps. To our knowledge, this is the first algorithm that enables single-pass alignment of multi-talker recordings. All algorithms are implemented using k2 / Icefall.

Dark humor often relies on subtle cultural nuances and implicit cues that require contextual reasoning to interpret, posing safety challenges that current static benchmarks fail to capture. To address this, we introduce a novel multimodal, multilingual benchmark for detecting and understanding harmful and offensive humor. Our manually curated dataset comprises 3,000 texts and 6,000 images in English and Arabic, alongside 1,200 videos that span English, Arabic, and language-independent (universal) contexts. Unlike standard toxicity datasets, we enforce a strict annotation guideline: distinguishing \emph{Safe} jokes from \emph{Harmful} ones, with the latter further classified into \emph{Explicit} (overt) and \emph{Implicit} (Covert) categories to probe deep reasoning. We systematically evaluate state-of-the-art (SOTA) open and closed-source models across all modalities. Our findings reveal that closed-source models significantly outperform open-source ones, with a notable difference in performance between the English and Arabic languages in both, underscoring the critical need for culturally grounded, reasoning-aware safety alignment. \textcolor{red}{Warning: this paper contains example data that may be offensive, harmful, or biased.}

Recent advances in autoregressive neural surrogate models have enabled orders-of-magnitude speedups in simulating dynamical systems. However, autoregressive models are generally prone to distribution drift: compounding errors in autoregressive rollouts that severely degrade generation quality over long time horizons. Existing work attempts to address this issue by implicitly leveraging the inherent trade-off between short-time accuracy and long-time consistency through hyperparameter tuning. In this work, we introduce a unifying mathematical framework that makes this tradeoff explicit, formalizing and generalizing hyperparameter-based strategies in existing approaches. Within this framework, we propose a robust, hyperparameter-free model implemented as a conditional diffusion model that balances short-time fidelity with long-time consistency by construction. Our model, Self-refining Neural Surrogate model (SNS), can be implemented as a standalone model that refines its own autoregressive outputs or as a complementary model to existing neural surrogates to ensure long-time consistency. We also demonstrate the numerical feasibility of SNS through high-fidelity simulations of complex dynamical systems over arbitrarily long time horizons.

Embeddings are a powerful way to enrich data-driven machine learning models with the world knowledge of large language models (LLMs). Yet, there is limited evidence on how to design effective LLM-based embedding pipelines for tabular prediction. In this work, we systematically benchmark 256 pipeline configurations, covering 8 preprocessing strategies, 16 embedding models, and 2 downstream models. Our results show that it strongly depends on the specific pipeline design whether incorporating the prior knowledge of LLMs improves the predictive performance. In general, concatenating embeddings tends to outperform replacing the original columns with embeddings. Larger embedding models tend to yield better results, while public leaderboard rankings and model popularity are poor performance indicators. Finally, gradient boosting decision trees tend to be strong downstream models. Our findings provide researchers and practitioners with guidance for building more effective embedding pipelines for tabular prediction tasks.

Recent advances in Large Vision-Language Models (LVLMs) have enabled training-free Fine-Grained Visual Recognition (FGVR). However, effectively exploiting LVLMs for FGVR remains challenging due to the inherent visual ambiguity of subordinate-level categories. Existing methods predominantly adopt either retrieval-oriented or reasoning-oriented paradigms to tackle this challenge, but both are constrained by two fundamental limitations:(1) They apply the same inference pipeline to all samples without accounting for uneven recognition difficulty, thereby leading to suboptimal accuracy and efficiency; (2) The lack of mechanisms to consolidate and reuse error-specific experience causes repeated failures on similar challenging cases. To address these limitations, we propose SARE, a Sample-wise Adaptive textbfREasoning framework for training-free FGVR. Specifically, SARE adopts a cascaded design that combines fast candidate retrieval with fine-grained reasoning, invoking the latter only when necessary. In the reasoning process, SARE incorporates a self-reflective experience mechanism that leverages past failures to provide transferable discriminative guidance during inference, without any parameter updates. Extensive experiments across 14 datasets substantiate that SARE achieves state-of-the-art performance while substantially reducing computational overhead.

Early prediction of Post-Acute Sequelae of SARS-CoV-2 severity is a critical challenge for women's health, particularly given the diagnostic overlap between PASC and common hormonal transitions such as menopause. Identifying and accounting for these confounding factors is essential for accurate long-term trajectory prediction. We conducted a retrospective study of 1,155 women (mean age 61) from the NIH RECOVER dataset. By integrating static clinical profiles with four weeks of longitudinal wearable data (monitoring cardiac activity and sleep), we developed a causal network based on a Large Language Model to predict future PASC scores. Our framework achieved a precision of 86.7\% in clinical severity prediction. Our causal attribution analysis demonstrate the model's ability to differentiate between active pathology and baseline noise: direct indicators such as breathlessness and malaise reached maximum saliency (1.00), while confounding factors like menopause and diabetes were successfully suppressed with saliency scores below 0.27.

Supervised detection of network attacks has always been a critical part of network intrusion detection systems (NIDS). Nowadays, in a pivotal time for artificial intelligence (AI), with even more sophisticated attacks that utilize advanced techniques, such as generative artificial intelligence (GenAI) and reinforcement learning, it has become a vital component if we wish to protect our personal data, which are scattered across the web. In this paper, we address two tasks, in the first unified multi-modal NIDS dataset, which incorporates flow-level data, packet payload information and temporal contextual features, from the reprocessed CIC-IDS-2017, CIC-IoT-2023, UNSW-NB15 and CIC-DDoS-2019, with the same feature space. In the first task we use machine learning (ML) algorithms, with stratified cross validation, in order to prevent network attacks, with stability and reliability. In the second task we use adversarial learning algorithms to generate synthetic data, compare them with the real ones and evaluate their fidelity, utility and privacy using the SDV framework, f-divergences, distinguishability and non-parametric statistical tests. The findings provide stable ML models for intrusion detection and generative models with high fidelity and utility, by combining the Synthetic Data Vault framework, the TRTS and TSTR tests, with non-parametric statistical tests and f-divergence measures.

Previous work has reported that vision foundation models show promising zero-shot performance in eye image segmentation. Here we examine whether the latest iteration of the Segment Anything Model, SAM3, offers better eye image segmentation performance than SAM2, and explore the performance of its new concept (text) prompting mode. Eye image segmentation performance was evaluated using diverse datasets encompassing both high-resolution high-quality videos from a lab environment and the TEyeD dataset consisting of challenging eye videos acquired in the wild. Results show that in most cases SAM3 with either visual or concept prompts did not perform better than SAM2, for both lab and in-the-wild datasets. Since SAM2 not only performed better but was also faster, we conclude that SAM2 remains the best option for eye image segmentation. We provide our adaptation of SAM3's codebase that allows processing videos of arbitrary duration.

Autonomous driving technologies have achieved significant advances in recent years, yet their real-world deployment remains constrained by data scarcity, safety requirements, and the need for generalization across diverse environments. In response, synthetic data and virtual environments have emerged as powerful enablers, offering scalable, controllable, and richly annotated scenarios for training and evaluation. This survey presents a comprehensive review of recent developments at the intersection of autonomous driving, simulation technologies, and synthetic datasets. We organize the landscape across three core dimensions: (i) the use of synthetic data for perception and planning, (ii) digital twin-based simulation for system validation, and (iii) domain adaptation strategies bridging synthetic and real-world data. We also highlight the role of vision-language models and simulation realism in enhancing scene understanding and generalization. A detailed taxonomy of datasets, tools, and simulation platforms is provided, alongside an analysis of trends in benchmark design. Finally, we discuss critical challenges and open research directions, including Sim2Real transfer, scalable safety validation, cooperative autonomy, and simulation-driven policy learning, that must be addressed to accelerate the path toward safe, generalizable, and globally deployable autonomous driving systems.

In the real economy, modern decision-making is fundamentally challenged by high-dimensional, multimodal environments, which are further complicated by agent heterogeneity and combinatorial data sparsity. This paper introduces a Multi-Agent Large Language Model-based Economic Sandbox (MALLES), leveraging the inherent generalization capabilities of large-sacle models to establish a unified simulation framework applicable to cross-domain and cross-category scenarios. Central to our approach is a preference learning paradigm in which LLMs are economically aligned via post-training on extensive, heterogeneous transaction records across diverse product categories. This methodology enables the models to internalize and transfer latent consumer preference patterns, thereby mitigating the data sparsity issues prevalent in individual categories. To enhance simulation stability, we implement a mean-field mechanism designed to model the dynamic interactions between the product environment and customer populations, effectively stabilizing sampling processes within high-dimensional decision spaces. Furthermore, we propose a multi-agent discussion framework wherein specialized agents collaboratively process extensive product information. This architecture distributes cognitive load to alleviate single-agent attention bottlenecks and captures critical decision factors through structured dialogue. Experiments demonstrate that our framework achieves significant improvements in product selection accuracy, purchase quantity prediction, and simulation stability compared to existing economic and financial LLM simulation baselines. Our results substantiate the potential of large language models as a foundational pillar for high-fidelity, scalable decision simulation and latter analysis in the real economy based on foundational database.

For LLM trading agents to be genuinely trustworthy, they must demonstrate understanding of market dynamics rather than exploitation of memorized ticker associations. Building responsible multi-agent systems demands rigorous signal validation: proving that predictions reflect legitimate patterns, not pre-trained recall. We address two sources of spurious performance: memorization bias from ticker-specific pre-training, and survivorship bias from flawed backtesting. Our approach is to blindfold the agents--anonymizing all identifiers--and verify whether meaningful signals persist. BlindTrade anonymizes tickers and company names, and four LLM agents output scores along with reasoning. We construct a GNN graph from reasoning embeddings and trade using PPO-DSR policy. On 2025 YTD (through 2025-08-01), we achieved Sharpe 1.40 +/- 0.22 across 20 seeds and validated signal legitimacy through negative control experiments. To assess robustness beyond a single OOS window, we additionally evaluate an extended period (2024--2025), revealing market-regime dependency: the policy excels in volatile conditions but shows reduced alpha in trending bull markets.

In this paper, we develop a stochastic set-valued optimization (SVO) framework tailored for robust machine learning. In the SVO setting, each decision variable is mapped to a set of objective values, and optimality is defined via set relations. We focus on SVO problems with hyperbox sets, which can be reformulated as multi-objective optimization (MOO) problems with finitely many objectives and serve as a foundation for representing or approximating more general mapped sets. Two special cases of hyperbox-valued optimization (HVO) are interval-valued (IVO) and rectangle-valued (RVO) optimization. We construct stochastic IVO/RVO formulations that incorporate subquantiles and superquantiles into the objective functions of the MOO reformulations, providing a new characterization for subquantiles. These formulations provide interpretable trade-offs by capturing both lower- and upper-tail behaviors of loss distributions, thereby going beyond standard empirical risk minimization and classical robust models. To solve the resulting multi-objective problems, we adopt stochastic multi-gradient algorithms and select a Pareto knee solution. In numerical experiments, the proposed algorithms with this selection strategy exhibit improved robustness and reduced variability across test replications under distributional shift compared with empirical risk minimization, while maintaining competitive accuracy.

We present a practical, reproducible framework for identifying undervalued football players grounded in objective mispricing. Instead of relying on subjective expert labels, we estimate an expected market value from structured data (historical market dynamics, biographical and contract features, transfer history) and compare it to the observed valuation to define mispricing. We then assess whether news-derived Natural Language Processing (NLP) features (i.e., sentiment statistics and semantic embeddings from football articles) complement market signals for shortlisting undervalued players. Using a chronological (leakage-aware) evaluation, gradient-boosted regression explains a large share of the variance in log-transformed market value. For undervaluation shortlisting, ROC-AUC-based ablations show that market dynamics are the primary signal, while NLP features provide consistent, secondary gains that improve robustness and interpretability. SHAP analyses suggest the dominance of market trends and age, with news-derived volatility cues amplifying signals in high-uncertainty regimes. The proposed pipeline is designed for decision support in scouting workflows, emphasizing ranking/shortlisting over hard classification thresholds, and includes a concise reproducibility and ethics statement.

Diffusion-based policies have gained significant popularity in Reinforcement Learning (RL) due to their ability to represent complex, non-Gaussian distributions. Stochastic Differential Equation (SDE)-based diffusion policies often rely on indirect entropy control due to the intractability of the exact entropy, while also suffering from computationally prohibitive policy gradients through the iterative denoising chain. To overcome these issues, we propose Flow Matching Policy with Entropy Regularization (FMER), an Ordinary Differential Equation (ODE)-based online RL framework. FMER parameterizes the policy via flow matching and samples actions along a straight probability path, motivated by optimal transport. FMER leverages the model's generative nature to construct an advantage-weighted target velocity field from a candidate set, steering policy updates toward high-value regions. By deriving a tractable entropy objective, FMER enables principled maximum-entropy optimization for enhanced exploration. Experiments on sparse multi-goal FrankaKitchen benchmarks demonstrate that FMER outperforms state-of-the-art methods, while remaining competitive on standard MuJoco benchmarks. Moreover, FMER reduces training time by 7x compared to heavy diffusion baselines (QVPO) and 10-15% relative to efficient variants.

Large language model (LLM) agents deployed in unknown environments must learn task structure at test time, but current approaches require thousands of interactions to form useful hypotheses. We present Sensi, an LLM agent architecture for the ARC-AGI-3 game-playing challenge that introduces structured test-time learning through three mechanisms: (1) a two-player architecture separating perception from action, (2) a curriculum-based learning system managed by an external state machine, and (3) a database-as-control-plane that makes the agents context window programmatically steerable. We further introduce an LLM-as-judge component with dynamically generated evaluation rubrics to determine when the agent has learned enough about one topic to advance to the next. We report results across two iterations: Sensi v1 solves 2 game levels using the two-player architecture alone, while Sensi v2 adds curriculum learning and solves 0 levels - but completes its entire learning curriculum in approximately 32 action attempts, achieving 50-94x greater sample efficiency than comparable systems that require 1600-3000 attempts. We precisely diagnose the failure mode as a self-consistent hallucination cascade originating in the perception layer, demonstrating that the architectural bottleneck has shifted from learning efficiency to perceptual grounding - a more tractable problem.

Existing vision-language models (VLMs) have demonstrated impressive performance in reasoning-based segmentation. However, current benchmarks are primarily constructed from high-quality images captured under idealized conditions. This raises a critical question: when visual cues are severely degraded by adverse weather conditions such as rain, snow, or fog, can VLMs sustain reliable reasoning segmentation capabilities? In response to this challenge, we introduce WeatherReasonSeg, a benchmark designed to evaluate VLM performance in reasoning-based segmentation under adverse weather conditions. It consists of two complementary components. First, we construct a controllable reasoning dataset by applying synthetic weather with varying severity levels to existing segmentation datasets, enabling fine-grained robustness analysis. Second, to capture real-world complexity, we curate a real-world adverse-weather reasoning segmentation dataset with semantically consistent queries generated via mask-guided LLM prompting. We further broaden the evaluation scope across five reasoning dimensions, including functionality, application scenarios, structural attributes, interactions, and requirement matching. Extensive experiments across diverse VLMs reveal two key findings: (1) VLM performance degrades monotonically with increasing weather severity, and (2) different weather types induce distinct vulnerability patterns. We hope WeatherReasonSeg will serve as a foundation for advancing robust, weather-aware reasoning.

Retrieval-Augmented Generation (RAG) improves factual grounding by incorporating external knowledge into language model generation. However, when retrieved context is noisy, unreliable, or inconsistent with the model's parametric knowledge, it introduces retrieval-prior conflicts that can degrade generation quality. While this problem has been studied in autoregressive language models, it remains largely unexplored in diffusion-based language models, where the iterative denoising process introduces unique challenges for integrating retrieved context. In this work, we propose Adaptive Retrieval-Augmented Masked Diffusion (ARAM), a training-free adaptive guidance framework for Masked Diffusion Models (MDMs) in RAG settings. ARAM dynamically calibrates the guidance scale during denoising according to the Signal-to-Noise Ratio (SNR) of the distributional shift induced by retrieved context. Intuitively, the model strengthens guidance when the retrieved context provides reliable corrective evidence and suppresses it when the contextual signal is noisy or non-supportive. Extensive experiments on multiple knowledge-intensive QA benchmarks show that ARAM improves overall QA performance over competitive RAG baselines.

Normalization is a critical operation in neural circuits. In the brain, there is evidence that normalization is implemented via inhibitory interneurons and allows neural populations to adjust to changes in the distribution of their inputs. In artificial neural networks (ANNs), normalization is used to improve learning in tasks that involve complex input distributions. However, it is unclear whether inhibition-mediated normalization in biological neural circuits also improves learning. Here, we explore this possibility using ANNs with separate excitatory and inhibitory populations trained on an image recognition task with variable luminosity. We find that inhibition-mediated normalization does not improve learning if normalization is applied only during inference. However, when this normalization is extended to include back-propagated errors, performance improves significantly. These results suggest that if inhibition-mediated normalization improves learning in the brain, it additionally requires the normalization of learning signals.

LLM agents are increasingly relevant to research domains such as vulnerability discovery. Yet, the strongest systems remain closed and cloud-only, making them resource-intensive, difficult to reproduce, and unsuitable for work involving proprietary code or sensitive data. Consequently, there is an urgent need for small, local models that can perform security tasks under strict resource budgets, but methods for developing them remain underexplored. In this paper, we address this gap by proposing a two-stage post-training pipeline. We focus on the problem of Linux privilege escalation, where success is automatically verifiable and the task requires multi-step interactive reasoning. Using an experimental setup that prevents data leakage, we post-train a 4B model in two stages: supervised fine-tuning on traces from procedurally generated privilege-escalation environments, followed by reinforcement learning with verifiable rewards. On a held-out benchmark of 12 Linux privilege-escalation scenarios, supervised fine-tuning alone more than doubles the baseline success rate at 20 rounds, and reinforcement learning further lifts our resulting model, PrivEsc-LLM, to 95.8%, nearly matching Claude Opus 4.6 at 97.5%. At the same time, the expected inference cost per successful escalation is reduced by over 100x.

Multimodal large language models (MLLMs) struggle with hallucinations, particularly with fine-grained queries, a challenge underrepresented by existing benchmarks that focus on coarse image-related questions. We introduce FIne-grained NEgative queRies (FINER), alongside two benchmarks: FINER-CompreCap and FINER-DOCCI. Using FINER, we analyze hallucinations across four settings: multi-object, multi-attribute, multi-relation, and ``what'' questions. Our benchmarks reveal that MLLMs hallucinate when fine-grained mismatches co-occur with genuinely present elements in the image. To address this, we propose FINER-Tuning, leveraging Direct Preference Optimization (DPO) on FINER-inspired data. Finetuning four frontier MLLMs with FINER-Tuning yields up to 24.2\% gains (InternVL3.5-14B) on hallucinations from our benchmarks, while simultaneously improving performance on eight existing hallucination suites, and enhancing general multimodal capabilities across six benchmarks. Code, benchmark, and models are available at \href{https://explainableml.github.io/finer-project/}{https://explainableml.github.io/finer-project/}.

The advent of Large Language Models (LLMs) represents a turning point in the theoretical foundations of Information Systems Engineering. Beyond their technical significance, LLMs challenge the ontological, epistemological, and semiotic assumptions that have long structured our understanding of in-formation, representation, and knowledge. This article proposes an integrative reflection on how LLMs reconfigure the relationships among language, meaning, and system design, suggesting that their emergence demands a re-examination of the conceptual foundations of contemporary information systems. Sketching on philosophical traditions from Peirce to Heidegger and Floridi, we investigate how the logic of generative models both extends and destabilises classical notions of ontology and signification. The discussion emphasises the necessity of grounding LLM-based systems in transparent, ethically coherent frameworks that respect the integrity of human-centred knowledge processes. Ultimately, the paper argues that LLMs should be understood not merely as tools for automation but as epistemic agents that reshape the philosophical and semiotic foundations of information systems engineering.

Cross-Domain Few-Shot Learning (CDFSL) adapts models trained with large-scale general data (source domain) to downstream target domains with only scarce training data, where the research on vision-language models (e.g., CLIP) is still in the early stages. Typical downstream domains, such as medical diagnosis, require fine-grained visual cues for interpretable recognition, but we find that current fine-tuned CLIP models can hardly focus on these cues, albeit they can roughly focus on important regions in source domains. Although current works have demonstrated CLIP's shortcomings in capturing local subtle patterns, in this paper, we find that the domain gap and scarce training data further exacerbate such shortcomings, much more than that of holistic patterns, which we call the local misalignment problem in CLIP-based CDFSL. To address this problem, due to the lack of supervision in aligning local visual features and text semantics, we turn to self-supervision information. Inspired by the translation task, we propose the CC-CDFSL method with cycle consistency, which translates local visual features into text features and then translates them back into visual features (and vice versa), and constrains the original features close to the translated back features. To reduce the noise imported by richer information in the visual modality, we further propose a Semantic Anchor mechanism, which first augments visual features to provide a larger corpus for the text-to-image mapping, and then shrinks the image features to filter out irrelevant image-to-text mapping. Extensive experiments on various benchmarks, backbones, and fine-tuning methods show we can (1) effectively improve the local vision-language alignment, (2) enhance the interpretability of learned patterns and model decisions by visualizing patches, and (3) achieve state-of-the-art performance.

Generative inbetweening (GI) seeks to synthesize realistic intermediate frames between the first and last keyframes beyond mere interpolation. As sequences become sparser and motions larger, previous GI models struggle with inconsistent frames with unstable pacing and semantic misalignment. Since GI involves fixed endpoints and numerous plausible paths, this task requires additional guidance gained from the keyframes and text to specify the intended path. Thus, we give semantic and temporal guidance from the keyframes and text onto each intermediate frame through Keyframe-anchored Attention Bias. We also better enforce frame consistency with Rescaled Temporal RoPE, which allows self-attention to attend to keyframes more faithfully. TGI-Bench, the first benchmark specifically designed for text-conditioned GI evaluation, enables challenge-targeted evaluation to analyze GI models. Without additional training, our method achieves state-of-the-art frame consistency, semantic fidelity, and pace stability for both short and long sequences across diverse challenges.

Requirements volatility is a major issue in software (SW) development, causing problems such as project delays and cost overruns. Even though there is a considerable amount of research related to requirement volatility, the majority of it is inclined toward project management aspects. The relationship between SW architecture design and requirements volatility has not been researched widely, even though changing requirements may for example lead to higher defect density during testing. An exploratory case study was conducted to study how requirements volatility affects SW architecture design. Fifteen semi-structured, thematic interviews were conducted in the case company, which provides the selection of software products for business customers and consumers. The research revealed the factors, such as requirements uncertainty and dynamic business environment, causing requirements volatility in the case company. The study identified the challenges that requirements volatility posed to SW architecture design, including scheduling and architectural technical debt. In addition, this study discusses means of mitigating the factors that cause requirements volatility and addressing the challenges posed by requirements volatility. SW architects are strongly influenced by requirement volatility. Thus understanding the factors causing requirements volatility as well as means to mitigate the challenges has high industrial relevance.

Although multigrid is asymptotically optimal for solving many important partial differential equations, its efficiency relies heavily on the careful selection of the individual algorithmic components. In contrast to recent approaches that can optimize certain multigrid components using deep learning techniques, we adopt a complementary strategy, employing evolutionary algorithms to construct efficient multigrid cycles from proven algorithmic building blocks. Here, we will present its application to generate efficient algebraic multigrid methods with so-called \emph{flexible cycling}, that is, level-specific smoothing sequences and non-recursive cycling patterns. The search space with such non-standard cycles is intractable to navigate manually, and is generated using genetic programming (GP) guided by context-free grammars. Numerical experiments with the linear algebra library, \emph{hypre}, demonstrate the potential of these non-standard GP cycles to improve multigrid performance both as a solver and a preconditioner.

Agentic AI has been a topic of great interest recently. A Large Language Model (LLM) agent involves one or more LLMs in the back-end. In the front end, it conducts autonomous decision-making by combining the LLM outputs with results obtained by invoking several external tools. The autonomous interactions with the external environment introduce critical security risks. In this paper, we present a grey-box approach to explore diverse behaviors and uncover security risks in LLM agents. Our approach VeriGrey uses the sequence of tools invoked as a feedback function to drive the testing process. This helps uncover infrequent but dangerous tool invocations that cause unexpected agent behavior. As mutation operators in the testing process, we mutate prompts to design pernicious injection prompts. This is carefully accomplished by linking the task of the agent to an injection task, so that the injection task becomes a necessary step of completing the agent functionality. Comparing our approach with a black-box baseline on the well-known AgentDojo benchmark, VeriGrey achieves 33% additional efficacy in finding indirect prompt injection vulnerabilities with a GPT-4.1 back-end. We also conduct real-world case studies with the widely used coding agent Gemini CLI, and the well-known OpenClaw personal assistant. VeriGrey finds prompts inducing several attack scenarios that could not be identified by black-box approaches. In OpenClaw, by constructing a conversation agent which employs mutational fuzz testing as needed, VeriGrey is able to discover malicious skill variants from 10 malicious skills (with 10/10= 100% success rate on the Kimi-K2.5 LLM backend, and 9/10= 90% success rate on Opus 4.6 LLM backend). This demonstrates the value of a dynamic approach like VeriGrey to test agents, and to eventually lead to an agent assurance framework.

We present DSS-GAN, the first generative adversarial network to employ Mamba as a hierarchical generator backbone for noise-to-image synthesis. The central contribution is Directional Latent Routing (DLR), a novel conditioning mechanism that decomposes the latent vector into direction-specific subvectors, each jointly projected with a class embedding to produce a feature-wise affine modulation of the corresponding Mamba scan. Unlike conventional class conditioning that injects a global signal, DLR couples class identity and latent structure along distinct spatial axes of the feature map, applied consistently across all generative scales. DSS-GAN achieves improved FID, KID, and precision-recall scores compared to StyleGAN2-ADA across multiple tested datasets. Analysis of the latent space reveals that directional subvectors exhibit measurable specialization: perturbations along individual components produce structured, direction-correlated changes in the synthesized image.

Understanding the dynamic behavior of biomolecules is fundamental to elucidating biological function and facilitating drug discovery. While Molecular Dynamics (MD) simulations provide a rigorous physical basis for studying these dynamics, they remain computationally expensive for long timescales. Conversely, recent deep generative models accelerate conformation generation but are typically either failing to model temporal relationship or built only for monomeric proteins. To bridge this gap, we introduce ATMOS, a novel generative framework based on State Space Models (SSM) designed to generate atom-level MD trajectories for biomolecular systems. ATMOS integrates a Pairformer-based state transition mechanism to capture long-range temporal dependencies, with a diffusion-based module to decode trajectory frames in an autoregressive manner. ATMOS is trained across crystal structures from PDB and conformation trajectory from large-scale MD simulation datasets including mdCATH and MISATO. We demonstrate that ATMOS achieves state-of-the-art performance in generating conformation trajectories for both protein monomers and complex protein-ligand systems. By enabling efficient inference of atomic trajectory of motions, this work establishes a promising foundation for modeling biomolecular dynamics.

The objective comparison of Reinforcement Learning (RL) algorithms is notoriously complex as outcomes and benchmarking of performances of different RL approaches are critically sensitive to environmental design, reward structures, and stochasticity inherent in both algorithmic learning and environmental dynamics. To manage this complexity, we introduce a rigorous benchmarking framework by extending converse optimality to discrete-time, control-affine, nonlinear systems with noise. Our framework provides necessary and sufficient conditions, under which a prescribed value function and policy are optimal for constructed systems, enabling the systematic generation of benchmark families via homotopy variations and randomized parameters. We validate it by automatically constructing diverse environments, demonstrating our framework's capacity for a controlled and comprehensive evaluation across algorithms. By assessing standard methods against a ground-truth optimum, our work delivers a reproducible foundation for precise and rigorous RL benchmarking.

Neural networks are central to modern artificial intelligence, yet their training remains highly sensitive to data contamination. Standard neural classifiers are trained by minimizing the categorical cross-entropy loss, corresponding to maximum likelihood estimation under a multinomial model. While statistically efficient under ideal conditions, this approach is highly vulnerable to contaminated observations including label noises corrupting supervision in the output space, and adversarial perturbations inducing worst-case deviations in the input space. In this paper, we propose a unified and statistically grounded framework for robust neural classification that addresses both forms of contamination within a single learning objective. We formulate neural network training as a minimum-divergence estimation problem and introduce rSDNet, a robust learning algorithm based on the general class of $S$-divergences. The resulting training objective inherits robustness properties from classical statistical estimation, automatically down-weighting aberrant observations through model probabilities. We establish essential population-level properties of rSDNet, including Fisher consistency, classification calibration implying Bayes optimality, and robustness guarantees under uniform label noise and infinitesimal feature contamination. Experiments on three benchmark image classification datasets show that rSDNet improves robustness to label corruption and adversarial attacks while maintaining competitive accuracy on clean data, Our results highlight minimum-divergence learning as a principled and effective framework for robust neural classification under heterogeneous data contamination.

Understanding whether large language models (LLMs) capture structured meaning requires examining how they represent concept relationships. In this work, we study three models of increasing scale: Pythia-70M, GPT-2, and Llama 3.1 8B, focusing on four semantic relations: synonymy, antonymy, hypernymy, and hyponymy. We combine linear probing with mechanistic interpretability techniques, including sparse autoencoders (SAE) and activation patching, to identify where these relations are encoded and how specific features contribute to their representation. Our results reveal a directional asymmetry in hierarchical relations: hypernymy is encoded redundantly and resists suppression, while hyponymy relies on compact features that are more easily disrupted by ablation. More broadly, relation signals are diffuse but exhibit stable profiles: they peak in the mid-layers and are stronger in post-residual/MLP pathways than in attention. Difficulty is consistent across models (antonymy easiest, synonymy hardest). Probe-level causality is capacity-dependent: on Llama 3.1, SAE-guided patching reliably shifts these signals, whereas on smaller models the shifts are weak or unstable. Our results clarify where and how reliably semantic relations are represented inside LLMs, and provide a reproducible framework for relating sparse features to probe-level causal evidence.

Federated Learning (FL) enables collaborative model training by sharing model updates instead of raw data, aiming to protect user privacy. However, recent studies reveal that these shared updates can inadvertently leak sensitive training data through gradient inversion attacks (GIAs). Among them, active GIAs are particularly powerful, enabling high-fidelity reconstruction of individual samples even under large batch sizes. Nevertheless, existing approaches often require architectural modifications, which limit their practical applicability. In this work, we bridge this gap by introducing the Activation REcovery via Sparse inversion (ARES) attack, an active GIA designed to reconstruct training samples from large training batches without requiring architectural modifications. Specifically, we formulate the recovery problem as a noisy sparse recovery task and solve it using the generalized Least Absolute Shrinkage and Selection Operator (Lasso). To extend the attack to multi-sample recovery, ARES incorporates the imprint method to disentangle activations, enabling scalable per-sample reconstruction. We further establish the expected recovery rate and derive an upper bound on the reconstruction error, providing theoretical guarantees for the ARES attack. Extensive experiments on CNNs and MLPs demonstrate that ARES achieves high-fidelity reconstruction across diverse datasets, significantly outperforming prior GIAs under large batch sizes and realistic FL settings. Our results highlight that intermediate activations pose a serious and underestimated privacy risk in FL, underscoring the urgent need for stronger defenses.

Reinforcement Learning (RL) has emerged as a powerful paradigm for training LLM-based agents, yet remains limited by low sample efficiency, stemming not only from sparse outcome feedback but also from the agent's inability to leverage prior experience across episodes. While augmenting agents with historical experience offers a promising remedy, existing approaches suffer from a critical weakness: the experience distilled from history is either stored statically or fail to coevolve with the improving actor, causing a progressive misalignment between the experience and the actor's evolving capability that diminishes its utility over the course of training. Inspired by complementary learning systems in neuroscience, we present Complementary RL to achieve seamless co-evolution of an experience extractor and a policy actor within the RL optimization loop. Specifically, the actor is optimized via sparse outcome-based rewards, while the experience extractor is optimized according to whether its distilled experiences demonstrably contribute to the actor's success, thereby evolving its experience management strategy in lockstep with the actor's growing capabilities. Empirically, Complementary RL outperforms outcome-based agentic RL baselines that do not learn from experience, achieving 10% performance improvement in single-task scenarios and exhibits robust scalability in multi-task settings. These results establish Complementary RL as a paradigm for efficient experience-driven agent learning.

Comprehending the information environment (IE) during crisis events is challenging due to the rapid change and abstract nature of the domain. Many approaches focus on snapshots via classification methods or network approaches to describe the IE in crisis, ignoring the temporal nature of how information changed over time. This work presents a system-oriented framework for modeling emerging narratives as temporally evolving semantic structures without requiring prior label specification. By integrating semantic embeddings, density-based clustering, and rolling temporal linkage, the framework represents narratives as persistent yet adaptive entities within a shared semantic space. We apply the methodology to a real-world crisis event and evaluate system behavior through stratified cluster validation and temporal lifecycle analysis. Results demonstrate high cluster coherence and reveal heterogeneous narrative lifecycles characterized by both transient fragments and stable narrative anchors. We ground our approach in situational awareness theory, supporting perception and comprehension of the IE by transforming unstructured social media streams into interpretable, temporally structured representations. The resulting system provides a methodology for monitoring and decision support in dynamic information environments.

LLMs have recently demonstrated strong capabilities in automatic RTL code generation, achieving high syntactic and functional correctness. However, most methods focus on functional correctness while overlooking critical physical design objectives, including Power, Performance, and Area. In this work, we propose a PPA-aware, tool-integrated multi-agent framework for high-quality verilog code generation. Our framework explicitly incorporates EDA tools into a closed-loop workflow composed of a \textit{Programmer Agent}, a \textit{Correctness Agent}, and a \textit{PPA Agent}, enabling joint optimization of functional correctness and physical metrics. To support continuous improvement without model retraining, we introduce an \textit{Evolved Memory Mechanism} that externalizes optimization experience into structured memory nodes. A dedicated memory manager dynamically maintains the memory pool and allows the system to refine strategies based on historical execution trajectories. Extensive experiments demonstrate that our approach achieves strong functional correctness while delivering significant improvements in PPA metrics. By integrating tool-driven feedback with structured and evolvable memory, our framework transforms RTL generation from one-shot reasoning into a continual, feedback-driven optimization process, providing a scalable pathway for deploying LLMs in real-world hardware design flows.

Multi-view learning primarily aims to fuse multiple features to describe data comprehensively. Most prior studies implicitly assume that different views share similar dimensions. In practice, however, severe dimensional disparities often exist among different views, leading to the unbalanced multi-view learning issue. For example, in emotion recognition tasks, video frames often reach dimensions of $10^6$, while physiological signals comprise only $10^1$ dimensions. Existing methods typically face two main challenges for this problem: (1) They often bias towards high-dimensional data, overlooking the low-dimensional views. (2) They struggle to effectively align representations under extreme dimensional imbalance, which introduces severe redundancy into the low-dimensional ones. To address these issues, we propose the Adaptive Multi-view Sparsity Learning (AdaMuS) framework. First, to prevent ignoring the information of low-dimensional views, we construct view-specific encoders to map them into a unified dimensional space. Given that mapping low-dimensional data to a high-dimensional space often causes severe overfitting, we design a parameter-free pruning method to adaptively remove redundant parameters in the encoders. Furthermore, we propose a sparse fusion paradigm that flexibly suppresses redundant dimensions and effectively aligns each view. Additionally, to learn representations with stronger generalization, we propose a self-supervised learning paradigm that obtains supervision information by constructing similarity graphs. Extensive evaluations on a synthetic toy dataset and seven real-world benchmarks demonstrate that AdaMuS consistently achieves superior performance and exhibits strong generalization across both classification and semantic segmentation tasks.

Climate change and the rapid growth of urban populations are intensifying environmental stresses within cities, making the behavior of urban atmospheric flows a critical factor in public health, energy use, and overall livability. This study targets to develop fast and accurate models of urban pollutant dispersion to support decision-makers, enabling them to implement mitigation measures in a timely and cost-effective manner. To reach this goal, an end-to-end data-driven approach is proposed to model and predict the airflow and pollutant dispersion in a street canyon in skimming flow regime. A series of time-resolved snapshots obtained from large eddy simulation (LES) serves as the database. The proposed framework is based on four fundamental steps. Firstly, a reduced basis is obtained by spectral proper orthogonal decomposition (SPOD) of the database. The projection of the time series snapshot data onto the SPOD modes (time-domain approach) provides the temporal coefficients of the dynamics. Secondly, a nonlinear compression of the temporal coefficients is performed by autoencoder to reduce further the dimensionality of the problem. Thirdly, a reduced-order model (ROM) is learned in the latent space using Long Short-Term Memory (LSTM) netowrks. Finally, the pollutant dispersion is estimated from the predicted velocity field through convolutional neural network that maps both fields. The results demonstrate the efficacy of the model in predicting the instantaneous as well as statistically stationary fields over long time horizon.

While context embeddings produced by LLMs can be used to estimate conceptual change, these representations are often not interpretable nor time-aware. Moreover, bias augmentation in historical data poses a non-trivial risk to researchers in the Digital Humanities. Hence, to model reliable concept trajectories in evolving scholarship, in this work we develop a framework that represents prototypical concepts through complex networks based on topics. Utilizing the Royal Society Corpus, we analyzed two competing theories from the Chemical Revolution (phlogiston vs. oxygen) as a case study to show that onomasiological change is linked to higher entropy and topological density, indicating increased diversity of ideas and connectivity effort.

This paper describes the design, implementation, and evaluation of a browser extension that provides contextual help to users who hover over technological acronyms and abbreviations on web pages. The extension combines a curated technical dictionary with OpenAI's large language model (LLM) to deliver on-demand definitions through lightweight tooltip overlays. A dual-layer artificial intelligence (AI) pipeline, comprising Google Cloud's Natural Language Processing (NLP) taxonomy API and OpenAI's ChatGPT, classifies each visited page as technology-related before activating the tooltip logic, thereby reducing false-positive detections. A mixed-methods study with 25 participants evaluated the tool's effect on reading comprehension and information-retrieval time among users with low to intermediate digital literacy. Results show that 92% of participants reported improved understanding of technical terms, 96% confirmed time savings over manual web searches, and all participants found the tooltips non-disruptive. Dictionary-based definitions were appended in an average of 2135 ms, compared to 16429 ms for AI-generated definitions and a mean manual search time of 17200 ms per acronym. The work demonstrates a practical, real-time approach to bridging the digital literacy gap and points toward extending contextual help to other domains such as medicine, law, and finance.

Large language models (LLMs) are currently applied to scientific paper evaluation by assigning an absolute score to each paper independently. However, since score scales vary across conferences, time periods, and evaluation criteria, models trained on absolute scores are prone to fitting narrow, context-specific rules rather than developing robust scholarly judgment. To overcome this limitation, we propose shifting paper evaluation from isolated scoring to collaborative ranking. In particular, we design \textbf{C}omparison-\textbf{N}ative framework for \textbf{P}aper \textbf{E}valuation (\textbf{CNPE}), integrating comparison into both data construction and model learning. We first propose a graph-based similarity ranking algorithm to facilitate the sampling of more informative and discriminative paper pairs from a collection. We then enhance relative quality judgment through supervised fine-tuning and reinforcement learning with comparison-based rewards. At inference, the model performs pairwise comparisons over sampled paper pairs and aggregates these preference signals into a global relative quality ranking. Experimental results demonstrate that our framework achieves an average relative improvement of \textbf{21.8\%} over the strong baseline DeepReview-14B, while exhibiting robust generalization to five previously unseen datasets. \href{https://github.com/ECNU-Text-Computing/ComparisonReview}{Code}.

Editing a 3D indoor scene from natural language is conceptually straightforward but technically challenging. Existing open-vocabulary systems often regenerate large portions of a scene or rely on image-space edits that disrupt spatial structure, resulting in unintended global changes or physically inconsistent layouts. These limitations stem from treating editing primarily as a generative task. We take a different view. A user instruction defines a desired world state, and editing should be the minimal sequence of actions that makes this state true while preserving everything else. This perspective motivates Edit-As-Act, a framework that performs open-vocabulary scene editing as goal-regressive planning in 3D space. Given a source scene and free-form instruction, Edit-As-Act predicts symbolic goal predicates and plans in EditLang, a PDDL-inspired action language that we design with explicit preconditions and effects encoding support, contact, collision, and other geometric relations. A language-driven planner proposes actions, and a validator enforces goal-directedness, monotonicity, and physical feasibility, producing interpretable and physically coherent transformations. By separating reasoning from low-level generation, Edit-As-Act achieves instruction fidelity, semantic consistency, and physical plausibility - three criteria that existing paradigms cannot satisfy together. On E2A-Bench, our benchmark of 63 editing tasks across 9 indoor environments, Edit-As-Act significantly outperforms prior approaches across all edit types and scene categories.

We present a drifting-based framework for amortized sampling of Boltzmann distributions defined by energy functions. The method trains a one-step neural generator by projecting samples along a Gaussian-smoothed score field from the current model distribution toward the target Boltzmann distribution. For targets specified only up to an unknown normalization constant, we derive a practical target-side drift from a smoothed energy and use two estimators: a local importance-sampling mean-shift estimator and a second-order curvature-corrected approximation. Combined with a mini-batch Gaussian mean-shift estimate of the sampler-side smoothed score, this yields a simple stop-gradient objective for stable one-step training. On a four-mode Gaussian-mixture Boltzmann target, our sampler achieves mean error $0.0754$, covariance error $0.0425$, and RBF MMD $0.0020$. Additional double-well and banana targets show that the same formulation also handles nonconvex and curved low-energy geometries. Overall, the results support drifting as an effective way to amortize iterative sampling from Boltzmann distributions into a single forward pass at test time.

Can latent actions and environment dynamics be recovered from offline trajectories when actions are never observed? We study this question in a setting where trajectories are action-free but tagged with demonstrator identity. We assume that each demonstrator follows a distinct policy, while the environment dynamics are shared across demonstrators and identity affects the next observation only through the chosen action. Under these assumptions, the conditional next-observation distribution $p(o_{t+1}\mid o_t,e)$ is a mixture of latent action-conditioned transition kernels with demonstrator-specific mixing weights. We show that this induces, for each state, a column-stochastic nonnegative matrix factorization of the observable conditional distribution. Using sufficiently scattered policy diversity and rank conditions, we prove that the latent transitions and demonstrator policies are identifiable up to permutation of the latent action labels. We extend the result to continuous observation spaces via a Gram-determinant minimum-volume criterion, and show that continuity of the transition map over a connected state space upgrades local permutation ambiguities to a single global permutation. A small amount of labeled action data then suffices to fix this final ambiguity. These results establish demonstrator diversity as a principled source of identifiability for learning latent actions and dynamics from offline RL data.

We propose SYRAN, an unsupervised anomaly detection method based on symbolic regression. Instead of encoding normal patterns in an opaque, high-dimensional model, our method learns an ensemble of human-readable equations that describe symbolic invariants: functions that are approximately constant on normal data. Deviations from these invariants yield anomaly scores, so that the detection logic is interpretable by construction, rather than via post-hoc explanation. Experimental results demonstrate that SYRAN is highly interpretable, providing equations that correspond to known scientific or medical relationships, and maintains strong anomaly detection performance comparable to that of state-of-the-art methods.

Vision-Language-Action (VLA) Models have become the mainstream solution for robot control, but suffer from slow inference speeds. Speculative Decoding (SD) is a promising acceleration method which can be divided into two categories: drafter-based SD and retrieval-based SD. Existing methods fail to analyze the advantages and disadvantages of these two types of SD in VLA models, leading to their sole application or optimization. In this paper, we analyze the trajectory patterns of robots controlled by the VLA model and derive a key insight: the two types of SD should be used in a hybrid manner. However, achieving hybrid SD in VLA models poses several challenges: (1) draft rejection and persistent errors in retrieval-based SD; (2) difficulty in determining the hybrid boundary. To address these, we propose the HeiSD framework. We propose a retrieval-based SD optimization method in HeiSD,which contains a verify-skip mechanism and a sequence-wise relaxed acceptance strategy. Moreover, we proposed a kinematic-based fused metric in HeiSD to automatically determine the hybrid boundary. Experimental results demonstrate that HeiSD attains a speedup of up to 2.45x in simulation benchmarks and 2.06x~2.41x in real-world scenarios, while sustaining a high task success rate.

Tabular foundation models, specifically Prior-Data Fitted Networks (PFNs), have revolutionized outlier detection (OD) by enabling unsupervised zero-shot adaptation to new datasets without training. However, despite their predictive power, these models typically function as opaque black boxes, outputting scalar outlier scores that lack the operational context required for safety-critical decision-making. Existing post-hoc explanation methods are often computationally prohibitive for real-time deployment or fail to capture the epistemic uncertainty inherent in zero-shot inference. In this work, we introduce FoMo-X, a modular framework that equips OD foundation models with intrinsic, lightweight diagnostic capabilities. We leverage the insight that the frozen embeddings of a pretrained PFN backbone already encode rich, context-conditioned relational information. FoMo-X attaches auxiliary diagnostic heads to these embeddings, trained offline using the same generative simulator prior as the backbone. This allows us to distill computationally expensive properties, such as Monte Carlo dropout based epistemic uncertainty, into a deterministic, single-pass inference. We instantiate FoMo-X with two novel heads: a Severity Head that discretizes deviations into interpretable risk tiers, and an Uncertainty Head that provides calibrated confidence measures. Extensive evaluation on synthetic and real-world benchmarks (ADBench) demonstrates that FoMo-X recovers ground-truth diagnostic signals with high fidelity and negligible inference overhead. By bridging the gap between foundation model performance and operational explainability, FoMo-X offers a scalable path toward trustworthy, zero-shot outlier detection.

Graph transformers are the state-of-the-art for learning from graph-structured data and are empirically known to avoid several pitfalls of message-passing architectures. However, there is limited theoretical analysis on why these models perform well in practice. In this work, we prove that attention-based architectures have structural benefits over graph convolutional networks in the context of node-level prediction tasks. Specifically, we study the neural network gaussian process limits of graph transformers (GAT, Graphormer, Specformer) with infinite width and infinite heads, and derive the node-level and edge-level kernels across the layers. Our results characterise how the node features and the graph structure propagate through the graph attention layers. As a specific example, we prove that graph transformers structurally preserve community information and maintain discriminative node representations even in deep layers, thereby preventing oversmoothing. We provide empirical evidence on synthetic and real-world graphs that validate our theoretical insights, such as integrating informative priors and positional encoding can improve performance of deep graph transformers.

Fine-tuning large language models (LLMs) with high-quality knowledge has been shown to enhance their performance effectively. However, there is a paucity of research on the depth of domain-specific knowledge comprehension by LLMs and the application of targeted active learning to improve their expertise. To address this gap, we introduce the Knowledge-Aware Active Learning (KA2L) framework. This framework assesses LLMs' mastery of specific knowledge points to aid in constructing unanswerable or unknowable questions through latent space analysis. This active learning strategy enhances training efficiency by focusing on knowledge the model has yet to master, thereby minimizing redundancy in learning already acquired information. This study innovatively employs a knowledge distribution probing technique to examine the hidden states of specific Transformer layers and identify the distribution of known and unknown knowledge within the LLM. Additionally, a hidden-state decoding method is proposed to generate numerous unknown questions in natural language from the latent knowledge space. In our experiments, we selected nine open-source LLMs to validate the effectiveness of the proposed framework. Results indicate that KA2L not only significantly reduces 50% annotation and computation costs across two open-domain and one vertical-domain dataset but also achieves better performance, offering valuable insights into active learning strategies for LLMs. The code is available at https://anonymous.4open.science/r/KA2L-F15C.

We introduce Emergent Trust Learning (ETL), a lightweight, trust-based control algorithm that can be plugged into existing AI agents. It enables these to reach cooperation in competitive game environments under shared resources. Each agent maintains a compact internal trust state, which modulates memory, exploration, and action selection. ETL requires only individual rewards and local observations and incurs negligible computational and communication overhead. We evaluate ETL in three environments: In a grid-based resource world, trust-based agents reduce conflicts and prevent long-term resource depletion while achieving competitive individual returns. In a hierarchical Tower environment with strong social dilemmas and randomised floor assignments, ETL sustains high survival rates and recovers cooperation even after extended phases of enforced greed. In the Iterated Prisoner's Dilemma, the algorithm generalises to a strategic meta-game, maintaining cooperation with reciprocal opponents while avoiding long-term exploitation by defectors. Code will be released upon publication.

Speech Large Language Models (Speech-LLMs) have emerged as a powerful approach for automatic speech recognition (ASR) by aligning speech encoders with large language models. However, adapting these systems to multilingual settings with imbalanced data distributions remains challenging. In such scenarios, a stability-plasticity dilemma often arises: fully shared Parameter-Efficient Fine-Tuning (PEFT) can cause negative inter-lingual interference for under-represented languages, while fully language-specific tuning limits the cross-lingual beneficial knowledge transfer needed for low-resource tasks. To address this, we propose Zipper-LoRA, a novel rank-level decoupling framework with three variants (Static, Hard, and Soft) that dynamically synthesizes LoRA updates from shared and language-specific subspaces. By using a lightweight language-conditioned router, Zipper-LoRA dynamically controls the contribution of each subspace at the LoRA rank level, enabling fine-grained sharing where languages are compatible and strict decoupling when conflicts occur. To further stabilize optimization under imbalanced data, we propose a two-stage training strategy with an Initial-B warm start that significantly accelerates convergence. Experiments on a 12-language mixed-resource setting show that Zipper-LoRA consistently outperforms both fully shared and independent baselines, particularly in extremely low-resource scenarios. Moreover, we demonstrate that these gains are robust across both chunked and non-chunked encoder configurations, confirming the framework's reliability for practical, large-scale multilingual ASR. Our code and data will be available at https://github.com/YuCeong-May/Zipper-LoRA for reproducibility.

Diffusion-based image-to-video (I2V) models are increasingly effective, yet they struggle to scale to ultra-high-resolution inputs (e.g., 4K). Generating videos at the model's native resolution often loses fine-grained structure, whereas high-resolution tiled denoising preserves local detail but breaks global layout consistency. This failure mode is particularly severe in the fresco animation setting: monumental artworks containing many distinct characters, objects, and semantically different sub-scenes that must remain spatially coherent over time. We introduce FrescoDiffusion, a training-free method for coherent large-format I2V generation from a single complex image. The key idea is to augment tiled denoising with a precomputed latent prior: we first generate a low-resolution video at the underlying model resolution and upsample its latent trajectory to obtain a global reference that captures long-range temporal and spatial structure. For 4K generation, we compute per-tile noise predictions and fuse them with this reference at every diffusion timestep by minimizing a single weighted least-squares objective in model-output space. The objective combines a standard tile-merging criterion with our regularization term, yielding a closed-form fusion update that strengthens global coherence while retaining fine detail. We additionally provide a spatial regularization variable that enables region-level control over where motion is allowed. Experiments on the VBench-I2V dataset and our proposed fresco I2V dataset show improved global consistency and fidelity over tiled baselines, while being computationally efficient. Our regularization enables explicit controllability of the trade-off between creativity and consistency.

We study the consistency of the $k$-nearest neighbor regressor under complex survey designs. While consistency results for this algorithm are well established for independent and identically distributed data, corresponding results for complex survey data are lacking. We show that the $k$-nearest neighbor regressor is consistent under regularity conditions on the sampling design and the distribution of the data. We derive lower bounds for the rate of convergence and show that these bounds exhibit the curse of dimensionality, as in the independent and identically distributed setting. Empirical studies based on simulated and real data illustrate our theoretical findings.

Estimating time-varying reproduction numbers from epidemic incidence data is a central task in infectious disease surveillance, yet it poses an inherently ill-posed inverse problem. Existing approaches often rely on strong structural assumptions derived from epidemiological models, which can limit their ability to adapt to non-stationary transmission dynamics induced by interventions or behavioral changes, leading to delayed detection of regime shifts and degraded estimation accuracy. In this work, we propose a Conditional Inverse Reproduction Learning framework (CIRL) that addresses the inverse problem by learning a {conditional mapping} from historical incidence patterns and explicit time information to latent reproduction numbers. Rather than imposing strongly enforced parametric constraints, CIRL softly integrates epidemiological structure with flexible likelihood-based statistical modeling, using the renewal equation as a forward operator to enforce dynamical consistency. The resulting framework combines epidemiologically grounded constraints with data-driven temporal representations, producing reproduction number estimates that are robust to observation noise while remaining responsive to abrupt transmission changes and zero-inflated incidence observations. Experiments on synthetic epidemics with controlled regime changes and real-world SARS and COVID-19 data demonstrate the effectiveness of the proposed approach.

Artificial intelligence systems predominantly rely on static data distributions, making them ineffective in dynamic real-world environments, such as cybersecurity, autonomous transportation, or finance, where data shifts frequently. Continual learning offers a potential solution by enabling models to learn from sequential data while retaining prior knowledge. However, a critical and underexplored issue in this domain is data normalization. Conventional normalization methods, such as min-max scaling, presuppose access to the entire dataset, which is incongruent with the sequential nature of continual learning. In this paper we introduce Continual Learning Adaptive Normalization (CLeAN), a novel adaptive normalization technique designed for continual learning in tabular data. CLeAN involves the estimation of global feature scales using learnable parameters that are updated via an Exponential Moving Average (EMA) module, enabling the model to adapt to evolving data distributions. Through comprehensive evaluations on two datasets and various continual learning strategies, including Resevoir Experience Replay, A-GEM, and EwC we demonstrate that CLeAN not only improves model performance on new data but also mitigates catastrophic forgetting. The findings underscore the importance of adaptive normalization in enhancing the stability and effectiveness of tabular data, offering a novel perspective on the use of normalization to preserve knowledge in dynamic learning environments.

Learning per-domain generalizing policies is a key challenge in learning for planning. Standard approaches learn state-value functions represented as graph neural networks using supervised learning on optimal plans generated by a teacher planner. In this work, we advocate for learning Q-value functions instead. Such policies are drastically cheaper to evaluate for a given state, as they need to process only the current state rather than every successor. Surprisingly, vanilla supervised learning of Q-values performs poorly as it does not learn to distinguish between the actions taken and those not taken by the teacher. We address this by using regularization terms that enforce this distinction, resulting in Q-value policies that consistently outperform state-value policies across a range of 10 domains and are competitive with the planner LAMA-first.

This paper outlines the conceptual and computational foundations of the AURORA (Acoustic Understanding and Real-time Observation of Resonant Articulations) model. AURORA predicts tongue displacement and shape in vowel sounds based on the first two formant values. It is intended as a didactic aid helping to explain the relationship between formants and the underlying articulation, as well as a foundation for biofeedback applications. The model is informed by ultrasound tongue imaging and acoustic data from 40 native speakers of English. In this paper we discuss the motivation for the model, the modelling objectives as well as the model architecture. We provide a qualitative evaluation of the model, focusing on selected tongue features. We then present two tools developed to make the model more accessible to a wider audience, a Shiny app and a prototype software for real-time tongue biofeedback. Potential users include students of phonetics, linguists in fields adjacent to phonetics, as well as speech and language therapy practitioners and clients.

Podcast listening is often grounded in a set of favorite shows, while listener intent can evolve over time. This combination of stable preferences and changing intent motivates recommendation approaches that support both familiarity and exploration. Traditional recommender systems typically emphasize long-term interaction patterns, and are less explicitly designed to incorporate rich contextual signals or flexible, intent-aware discovery objectives. In this setting, models that can jointly reason over semantics, context, and user state offer a promising direction. Large Language Models (LLMs) provide strong semantic reasoning and contextual conditioning for discovery-oriented recommendation, but deploying them in production introduces challenges in catalog grounding, user-level personalization, and latency-critical serving. We address these challenges with GLIDE, a production-scale generative recommender for podcast discovery at Spotify. GLIDE formulates recommendation as an instruction-following task over a discretized catalog using Semantic IDs, enabling grounded generation over a large inventory. The model conditions on recent listening history and lightweight user context, while injecting long-term user embeddings as soft prompts to capture stable preferences under strict inference constraints. We evaluate GLIDE using offline retrieval metrics, human judgments, and LLM-based evaluation, and validate its impact through large-scale online A/B testing. Across experiments involving millions of users, GLIDE increases non-habitual podcast streaming on Spotify home surface by up to 5.4% and new-show discovery by up to 14.3%, while meeting production cost and latency constraints.

A symmetry on rigid motion is one of the salient factors in efficient learning of 3D point cloud problems. Group convolution has been a representative method to extract equivariant features, but its realizations have struggled to retain both rigorous symmetry and scalability simultaneously. We advocate utilizing the intertwiner framework to resolve this trade-off, but previous works on it, which did not achieve complete SE(3) symmetry or scalability to large-scale problems, necessitate a more advanced kernel architecture. We present Equivariant Coordinate-based Kernel Convolution, or ECKConv. It acquires SE(3) equivariance from the kernel domain defined in a double coset space, and its explicit kernel design using coordinate-based networks enhances its learning capability and memory efficiency. The experiments on diverse point cloud tasks, e.g., classification, pose registration, part segmentation, and large-scale semantic segmentation, validate the rigid equivariance, memory scalability, and outstanding performance of ECKConv compared to state-of-the-art equivariant methods.

In many CAD-based applications, complex geometries are defined by a high number of design parameters. This leads to high-dimensional design spaces that are challenging for downstream engineering processes like simulations, optimization, and design exploration tasks. Therefore, dimension reduction methods such as principal component analysis (PCA) are used. The PCA identifies dominant modes of geometric variation and yields a compact representation of the geometry. While classical PCA excels in the compact representation part, it does not directly recover underlying design parameters of a generated geometry. In this work, we deal with the problem of estimating design parameters from PCA-based representations. Analyzing a recent modification of the PCA dedicated to our field of application, we show that the results are actually identical to the standard PCA. We investigate limitations of this approach and present reasonable conditions under which accurate, interpretable parameter estimation can be obtained. With the help of dedicated experiments, we take a more in-depth look at every stage of the PCA and the possible changes of the geometry during these processes.

Recently, in eXplainable AI (XAI), $\textit{even if}$ explanations -- so-called semi-factuals -- have emerged as a popular strategy that explains how a predicted outcome $\textit{can remain the same}$ even when certain input-features are altered. For example, in the commonly-used banking app scenario, a semi-factual explanation could inform customers about better options, other alternatives for their successful application, by saying "$\textit{Even if}$ you asked for double the loan amount, you would still be accepted". Most semi-factuals XAI algorithms focus on finding maximal value-changes to a single key-feature that do $\textit{not}$ alter the outcome (unlike counterfactual explanations that often find minimal value-changes to several features that alter the outcome). However, no current semi-factual method explains $\textit{why}$ these extreme value-changes do not alter outcomes; for example, a more informative semi-factual could tell the customer that it is their good credit score that allows them to borrow double their requested loan. In this work, we advance a new algorithm -- the $\textit{informative semi-factuals}$ (ISF) method -- that generates more elaborated explanations supplementing semi-factuals with information about additional $\textit{hidden features}$ that influence an automated decision. Experimental results on benchmark datasets show that this ISF method computes semi-factuals that are both informative and of high-quality on key metrics. Furthermore, a user study shows that people prefer these elaborated explanations over the simpler semi-factual explanations generated by current methods.

LLMs are increasingly applied to recommendation, retrieval, and reasoning, yet deploying a single end-to-end model that can jointly support these behaviors over large, heterogeneous catalogs remains challenging. Such systems must generate unambiguous references to real items, handle multiple entity types, and operate under strict latency and reliability constraints requirements that are difficult to satisfy with text-only generation. While tool-augmented recommender systems address parts of this problem, they introduce orchestration complexity and limit end-to-end optimization. We view this setting as an instance of a broader research problem: how to adapt LLMs to reason jointly over multiple-domain entities, users, and language in a fully self-contained manner. To this end, we introduce NEO, a framework that adapts a pre-trained decoder-only LLM into a tool-free, catalog-grounded generator. NEO represents items as SIDs and trains a single model to interleave natural language and typed item identifiers within a shared sequence. Text prompts control the task, target entity type, and output format (IDs, text, or mixed), while constrained decoding guarantees catalog-valid item generation without restricting free-form text. We refer to this instruction-conditioned controllability as language-steerability. We treat SIDs as a distinct modality and study design choices for integrating discrete entity representations into LLMs via staged alignment and instruction tuning. We evaluate NEO at scale on a real-world catalog of over 10M items across multiple media types and discovery tasks, including recommendation, search, and user understanding. In offline experiments, NEO consistently outperforms strong task-specific baselines and exhibits cross-task transfer, demonstrating a practical path toward consolidating large-scale discovery capabilities into a single language-steerable generative model.

Accurate prediction of permeability tensors from pore-scale microstructure images is essential for subsurface flow modeling, yet direct numerical simulation requires hours per sample, fundamentally limiting large-scale uncertainty quantification and reservoir optimization workflows. A physics-informed deep learning framework is presented that resolves this bottleneck by combining a MaxViT hybrid CNN-Transformer architecture with progressive transfer learning and differentiable physical constraints. MaxViT's multi-axis attention mechanism simultaneously resolves grain-scale pore-throat geometry via block-local operations and REV-scale connectivity statistics through grid-global operations, providing the spatial hierarchy that permeability tensor prediction physically requires. Training on 20000 synthetic porous media samples spanning three orders of magnitude in permeability, a three-phase progressive curriculum advances from an ImageNet-pretrained baseline with D4-equivariant augmentation and tensor transformation, through component-weighted loss prioritizing off-diagonal coupling, to frozen-backbone transfer learning with porosity conditioning via Feature-wise Linear Modulation (FiLM). Onsager reciprocity and positive definiteness are enforced via differentiable penalty terms. On a held-out test set of 4000 samples, the framework achieves variance-weighted R2 = 0.9960 (R2_Kxx = 0.9967, R2_Kxy = 0.9758), a 33% reduction in unexplained variance over the supervised baseline. The results offer three transferable principles for physics-informed scientific machine learning: large-scale visual pretraining transfers effectively across domain boundaries; physical constraints are most robustly integrated as differentiable architectural components; and progressive training guided by diagnostic failure-mode analysis enables unambiguous attribution of performance gains across methodological stages.

Recent advancements in diffusion-based image editing pose a significant threat to the authenticity of digital visual content. Traditional embedding-based watermarking methods often introduce perceptible perturbations to maintain robustness, inevitably compromising visual fidelity. Meanwhile, existing zero-watermarking approaches, typically relying on global image features, struggle to withstand sophisticated manipulations. In this work, we uncover a key observation: while individual image patches undergo substantial alterations during AI-based editing, the relational distance between patch pairs remains relatively invariant. Leveraging this property, we propose Relational Zero-Watermarking (Rel-Zero), a novel framework that requires no modification to the original image but derives a unique zero-watermark from these editing-invariant patch relations. By grounding the watermark in intrinsic structural consistency rather than absolute appearance, Rel-Zero provides a non-invasive yet resilient mechanism for content authentication. Extensive experiments demonstrate that Rel-Zero achieves substantially improved robustness across diverse editing models and manipulations compared to prior zero-watermarking approaches.

Visual Anomaly Detection (VAD) is crucial for industrial inspection, yet most existing methods are limited to single-category scenarios, failing to address the multi-class and continual learning demands of real-world environments. While Teacher-Student (TS) architectures are efficient, they remain unexplored for the Continual Setting. To bridge this gap, we propose AdapTS, a unified TS framework designed for multi-class and continual settings, optimized for edge deployment. AdapTS eliminates the need for two different architectures by utilizing a single shared frozen backbone and injecting lightweight trainable adapters into the student pathway. Training is enhanced via a segmentation-guided objective and synthetic Perlin noise, while a prototype-based task identification mechanism dynamically selects adapters at inference with 99\% accuracy. Experiments on MVTec AD and VisA demonstrate that AdapTS matches the performance of existing TS methods across multi-class and continual learning scenarios, while drastically reducing memory overhead. Our lightest variant, AdapTS-S, requires only 8 MB of additional memory, 13x less than STFPM (95 MB), 48x less than RD4AD (360 MB), and 149x less than DeSTSeg (1120 MB), making it a highly scalable solution for edge deployment in complex industrial environments.

Accurate air quality forecasting is essential for public health and environmental sustainability, but remains challenging due to the complex pollutant dynamics. Existing deep learning methods often model pollutant dynamics as an instantaneous process, overlooking the intrinsic delays in pollutant propagation. Thus, we propose AirDDE, the first neural delay differential equation framework in this task that integrates delay modeling into a continuous-time pollutant evolution under physical guidance. Specifically, two novel components are introduced: (1) a memory-augmented attention module that retrieves globally and locally historical features, which can adaptively capture delay effects modulated by multifactor data; and (2) a physics-guided delay evolving function, grounded in the diffusion-advection equation, that models diffusion, delayed advection, and source/sink terms, which can capture delay-aware pollutant accumulation patterns with physical plausibility. Extensive experiments on three real-world datasets demonstrate that AirDDE achieves the state-of-the-art forecasting performance with an average MAE reduction of 8.79\% over the best baselines. The code is available at https://github.com/w2obin/airdde-aaai.

Mirror Descent (MD) is a scalable first-order method widely used in large-scale optimization, with applications in image processing, policy optimization, and neural network training. This paper generalizes MD to optimization on Riemannian manifolds. In particular, we develop a Riemannian Mirror Descent (RMD) framework via reparameterization and further propose a stochastic variant of RMD. We also establish non-asymptotic convergence guarantees for both RMD and stochastic RMD. As an application to the Stiefel manifold, our RMD framework reduces to the Curvilinear Gradient Descent (CGD) method proposed in [26]. Moreover, when specializing the stochastic RMD framework to the Stiefel setting, we obtain a stochastic extension of CGD, which effectively addresses large-scale manifold optimization problems.

In this paper, we introduce a novel kinematics-rich vision-language-action (VLA) task, in which language commands densely encode diverse kinematic attributes (such as direction, trajectory, orientation, and relative displacement) from initiation through completion, at key moments, unlike existing action instructions that capture kinematics only coarsely or partially, thereby supporting fine-grained and personalized manipulation. In this setting, where task goals remain invariant while execution trajectories must adapt to instruction-level kinematic specifications. To address this challenge, we propose KineVLA, a vision-language-action framework that explicitly decouples goal-level invariance from kinematics-level variability through a bi-level action representation and bi-level reasoning tokens to serve as explicit, supervised intermediate variables that align language and action. To support this task, we construct the kinematics-aware VLA datasets spanning both simulation and real-world robotic platforms, featuring instruction-level kinematic variations and bi-level annotations. Extensive experiments on LIBERO and a Realman-75 robot demonstrate that KineVLA consistently outperforms strong VLA baselines on kinematics-sensitive benchmarks, achieving more precise, controllable, and generalizable manipulation behaviors.

Neural Operators (NOs) are a powerful deep learning framework designed to learn the solution operator that arise from partial differential equations. This study investigates NOs ability to capture the stiff spatio-temporal dynamics of the FitzHugh-Nagumo model, which describes excitable cells. A key contribution of this work is evaluating the translation invariance using a novel training strategy. NOs are trained using an applied current with varying spatial locations and intensities at a fixed time, and the test set introduces a more challenging out-of-distribution scenario in which the applied current is translated in both time and space. This approach significantly reduces the computational cost of dataset generation. Moreover we benchmark seven NOs architectures: Convolutional Neural Operators (CNOs), Deep Operator Networks (DONs), DONs with CNN encoder (DONs-CNN), Proper Orthogonal Decomposition DONs (POD-DONs), Fourier Neural Operators (FNOs), Tucker Tensorized FNOs (TFNOs), Localized Neural Operators (LocalNOs). We evaluated these models based on training and test accuracy, efficiency, and inference speed. Our results reveal that CNOs performs well on translated test dynamics. However, they require higher training costs, though their performance on the training set is similar to that of the other considered architectures. In contrast, FNOs achieve the lowest training error, but have the highest inference time. Regarding the translated dynamics, FNOs and their variants provide less accurate predictions. Finally, DONs and their variants demonstrate high efficiency in both training and inference, however they do not generalize well to the test set. These findings highlight the current capabilities and limitations of NOs in capturing complex ionic model dynamics and provide a comprehensive benchmark including their application to scenarios involving translated dynamics.

The rapid proliferation of large language models (LLMs) has created an urgent need for robust and generalizable detectors of machine-generated text. Existing benchmarks typically evaluate a single detector on a single dataset under ideal conditions, leaving open questions about cross-domain transfer, cross-LLM generalization, and adversarial robustness. We present a comprehensive benchmark evaluating diverse detection approaches across two corpora: HC3 (23,363 human-ChatGPT pairs) and ELI5 (15,000 human-Mistral-7B pairs). Methods include classical classifiers, fine-tuned transformer encoders (BERT, RoBERTa, ELECTRA, DistilBERT, DeBERTa-v3), a CNN, an XGBoost stylometric model, perplexity-based detectors, and LLM-as-detector prompting. Results show that transformer models achieve near-perfect in-distribution performance but degrade under domain shift. The XGBoost stylometric model matches performance while remaining interpretable. LLM-based detectors underperform and are affected by generator-detector identity bias. Perplexity-based methods exhibit polarity inversion, with modern LLM outputs showing lower perplexity than human text, but remain effective when corrected. No method generalizes robustly across domains and LLM sources.

Large language models (LLMs) exhibit strong general intelligence, yet their multilingual performance remains highly imbalanced. Although LLMs encode substantial cross-lingual knowledge in a unified semantic space, they often struggle to reliably interface this knowledge with low-resource or unseen languages. Fortunately, pretrained encoder-decoder translation models already possess balanced multilingual capability, suggesting a natural complement to LLMs. In this work, we propose XBridge, a compositional encoder-LLM-decoder architecture that offloads multilingual understanding and generation to external pretrained translation models, while preserving the LLM as an English-centric core for general knowledge processing. To address the resulting representation misalignment across models, we introduce lightweight cross-model mapping layers and an optimal transport-based alignment objective, enabling fine-grained semantic consistency for multilingual generation. Experiments on four LLMs across multilingual understanding, reasoning, summarization, and generation indicate that XBridge outperforms strong baselines, especially on low-resource and previously unseen languages, without retraining the LLM.

Federated Learning (FL) enables privacy-preserving intelligence on Internet of Things (IoT) devices but incurs a significant carbon footprint due to the high energy cost of frequent uplink transmission. While pre-trained models are increasingly available on edge devices, their potential to reduce the energy overhead of fine-tuning remains underexplored. In this work, we propose QuantFL, a sustainable FL framework that leverages pre-trained initialisation to enable aggressive, computationally lightweight quantisation. We demonstrate that pre-training naturally concentrates update statistics, allowing us to use memory-efficient bucket quantisation without the energy-intensive overhead of complex error-feedback mechanisms. On MNIST and CIFAR-100, QuantFL reduces total communication by 40\% ($\simeq40\%$ total-bit reduction with full-precision downlink; $\geq80\%$ on uplink or when downlink is quantised) while matching or exceeding uncompressed baselines under strict bandwidth budgets; BU attains 89.00\% (MNIST) and 66.89\% (CIFAR-100) test accuracy with orders of magnitude fewer bits. We also account for uplink and downlink costs and provide ablations on quantisation levels and initialisation. QuantFL delivers a practical, "green" recipe for scalable training on battery-constrained IoT networks.

Large language models (LLMs) often hallucinate, producing fluent but false information, partly because supervised fine-tuning (SFT) implicitly rewards always responding. We introduce $\textit{HypoTermInstruct}$, an SFT dataset (31,487 responses for 11,151 questions) designed to teach models epistemological humility-the ability to recognize the limits of their own knowledge and admit uncertainty. This is achieved through questions about non-existent "hypothetical" terms. We also release $\textit{HypoTermQA-Enhanced}$, a benchmark for hallucination tendency strengthened through multiple validations. We conducted 800 controlled LoRA SFT runs across $\textit{Llama3.1-8B}$ and $\textit{Gemma3-4B}$ (base and instruct), testing 100 fine-tuning configurations with paired controls. Our results demonstrate that replacing generic instruction data with $\textit{HypoTermInstruct}$ significantly improves the HypoTerm Score (median increases of 0.19% to 25.91%) and FactScore (+0.39% to +0.86%), while maintaining stable performance on MMLU (minimal decreases of 0.26% to 0.35%). Our work demonstrates that targeted, high-quality SFT data teaching meta-cognitive skills can effectively reduce hallucination without preference/RL pipelines, providing mechanistic insights and a practical path toward more reliable AI systems.

Language models struggle to generalize beyond pretraining context lengths, limiting long-horizon reasoning and retrieval. Continued pretraining on long-context data can help but is expensive due to the quadratic scaling of Attention. We observe that most tokens do not require (Global) Attention over the entire sequence and can rely on local context. Based on this, we propose L2A (Learning To Attend), a layer that enables conditional (token-wise) long-range memory access by deciding when to invoke global attention. We evaluate L2A on Qwen 2.5 and Qwen 3 models, extending their effective context length from 32K to 128K tokens. L2A matches the performance of standard long-context training to within 3% while skipping Global Attention for $\sim$80% of tokens, outperforming prior baselines. We also design custom Triton kernels to efficiently implement this token-wise conditional Attention on GPUs, achieving up to $\sim$2x improvements in training throughput and time-to-first-token over FlashAttention. Moreover, L2A enables post-training pruning of highly sparse Global Attention layers, reducing KV cache memory by up to 50% with negligible performance loss.

This study explores the combination of automated machine learning (AutoML) with model-based deep unfolding (DU) for optimizing wireless beamforming and waveforms. We convert the iterative proximal gradient descent (PGD) algorithm into a deep neural network, wherein the parameters of each layer are learned instead of being predetermined. Additionally, we enhance the architecture by incorporating a hybrid layer that performs a learnable linear gradient transformation prior to the proximal projection. By utilizing AutoGluon with a tree-structured parzen estimator (TPE) for hyperparameter optimization (HPO) across an expanded search space, which includes network depth, step-size initialization, optimizer, learning rate scheduler, layer type, and post-gradient activation, the proposed auto-unrolled PGD (Auto-PGD) achieves 98.8% of the spectral efficiency of a traditional 200-iteration PGD solver using only five unrolled layers, while requiring only 100 training samples. We also address a gradient normalization issue to ensure consistent performance during training and evaluation, and we illustrate per-layer sum-rate logging as a tool for transparency. These contributions highlight a notable reduction in the amount of training data and inference cost required, while maintaining high interpretability compared to conventional black-box architectures.

Unified Multimodal Models (UMMs) offer powerful cross-modality capabilities but introduce new safety risks not observed in single-task models. Despite their emergence, existing safety benchmarks remain fragmented across tasks and modalities, limiting the comprehensive evaluation of complex system-level vulnerabilities. To address this gap, we introduce UniSAFE, the first comprehensive benchmark for system-level safety evaluation of UMMs across 7 I/O modality combinations, spanning conventional tasks and novel multimodal-context image generation settings. UniSAFE is built with a shared-target design that projects common risk scenarios across task-specific I/O configurations, enabling controlled cross-task comparisons of safety failures. Comprising 6,802 curated instances, we use UniSAFE to evaluate 15 state-of-the-art UMMs, both proprietary and open-source. Our results reveal critical vulnerabilities across current UMMs, including elevated safety violations in multi-image composition and multi-turn settings, with image-output tasks consistently more vulnerable than text-output tasks. These findings highlight the need for stronger system-level safety alignment for UMMs. Our code and data are publicly available at https://github.com/segyulee/UniSAFE

Categorization is a core component of human linguistic competence. We investigate how a transformer-based language model (LM) learns linguistic categories by comparing its behaviour over the course of training to behaviours which characterize abstract feature-based and concrete exemplar-based accounts of human language acquisition. We investigate how lexical semantic and syntactic categories emerge using novel divergence-based metrics that track learning trajectories using next-token distributions. In experiments with GPT-2 small, we find that (i) when a construction is learned, abstract class-level behaviour is evident at earlier steps than lexical item-specific behaviour, and (ii) that different linguistic behaviours emerge abruptly in sequence at different points in training, revealing that abstraction plays a key role in how LMs learn. This result informs the models of human language acquisition that LMs may serve as an existence proof for.

Unsupervised Domain Adaptation (UDA) seeks to transfer knowledge from a labeled source domain to an unlabeled target domain but often suffers from severe domain and scale gaps that degrade performance. Existing cross-attention-based transformers can align features across domains, yet they struggle to preserve content semantics under large appearance and scale variations. To explicitly address these challenges, we introduce the concept of beneficial noise, which regularizes cross-attention by injecting controlled perturbations, encouraging the model to ignore style distractions and focus on content. We propose the Domain-Adaptive Cross-Scale Matching (DACSM) framework, which consists of a Domain-Adaptive Transformer (DAT) for disentangling domain-shared content from domain-specific style, and a Cross-Scale Matching (CSM) module that adaptively aligns features across multiple resolutions. DAT incorporates beneficial noise into cross-attention, enabling progressive domain translation with enhanced robustness, yielding content-consistent and style-invariant representations. Meanwhile, CSM ensures semantic consistency under scale changes. Extensive experiments on VisDA-2017, Office-Home, and DomainNet demonstrate that DACSM achieves state-of-the-art performance, with up to +2.3% improvement over CDTrans on VisDA-2017. Notably, DACSM achieves a +5.9% gain on the challenging "truck" class of VisDA, evidencing the strength of beneficial noise in handling scale discrepancies. These results highlight the effectiveness of combining domain translation, beneficial-noise-enhanced attention, and scale-aware alignment for robust cross-domain representation learning.

Monocular 3D object detection typically relies on pseudo-labeling techniques to reduce dependency on real-world annotations. Recent advances demonstrate that deterministic linguistic cues can serve as effective auxiliary weak supervision signals, providing complementary semantic context. However, hand-crafted textual descriptions struggle to capture the inherent visual diversity of individuals across scenes, limiting the model's ability to learn scene-aware representations. To address this challenge, we propose Visual-referred Probabilistic Prompt Learning (VirPro), an adaptive multi-modal pretraining paradigm that can be seamlessly integrated into diverse weakly supervised monocular 3D detection frameworks. Specifically, we generate a diverse set of learnable, instance-conditioned prompts across scenes and store them in an Adaptive Prompt Bank (APB). Subsequently, we introduce Multi-Gaussian Prompt Modeling (MGPM), which incorporates scene-based visual features into the corresponding textual embeddings, allowing the text prompts to express visual uncertainties. Then, from the fused vision-language embeddings, we decode a prompt-targeted Gaussian, from which we derive a unified object-level prompt embedding for each instance. RoI-level contrastive matching is employed to enforce modality alignment, bringing embeddings of co-occurring objects within the same scene closer in the latent space, thus enhancing semantic coherence. Extensive experiments on the KITTI benchmark demonstrate that integrating our pretraining paradigm consistently yields substantial performance gains, achieving up to a 4.8% average precision improvement than the baseline.

We present GuidedSAC, a novel reinforcement learning (RL) algorithm that facilitates efficient exploration in vast state-action spaces. GuidedSAC leverages large language models (LLMs) as intelligent supervisors that provide action-level guidance for the Soft Actor-Critic (SAC) algorithm. The LLM-based supervisor analyzes the most recent trajectory using state information and visual replays, offering action-level interventions that enable targeted exploration. Furthermore, we provide a theoretical analysis of GuidedSAC, proving that it preserves the convergence guarantees of SAC while improving convergence speed. Through experiments in both discrete and continuous control environments, including toy text tasks and complex MuJoCo benchmarks, we demonstrate that GuidedSAC consistently outperforms standard SAC and state-of-the-art exploration-enhanced variants (e.g., RND, ICM, and E3B) in terms of sample efficiency and final performance.

Sequential Recommendation (SR) in multimodal settings typically relies on small frozen pretrained encoders, which limits semantic capacity and prevents Collaborative Filtering (CF) signals from being fully integrated into item representations. Inspired by the recent success of Large Language Models (LLMs) as high-capacity embedders, we investigate the use of Vision-Language Models (VLMs) as CF-aware multimodal encoders for SR. However, we find that standard contrastive supervised fine-tuning (SFT), which adapts VLMs for embedding generation and injects CF signals, can amplify its inherent modality collapse. In this state, optimization is dominated by a single modality while the other degrades, ultimately undermining recommendation accuracy. To address this, we propose VLM2Rec, a VLM embedder-based framework for multimodal sequential recommendation designed to ensure balanced modality utilization. Specifically, we introduce Weak-modality Penalized Contrastive Learning to rectify gradient imbalance during optimization and Cross-Modal Relational Topology Regularization to preserve geometric consistency between modalities. Extensive experiments demonstrate that VLM2Rec consistently outperforms state-of-the-art baselines in both accuracy and robustness across diverse scenarios.

Large language models achieve breakthroughs in complex reasoning via long chain-of-thought sequences. However, this often leads to severe reasoning inflation, causing substantial computational redundancy. To maximize Intelligence per Token, we introduce a theoretical metric, MSL-Minimal Sufficient Length. MSL rigorously characterizes the shortest reasoning length that preserves answer correctness. We provide a recursive definition based on independently sampled sequences and prove the existence of its limit, establishing the first measurable lower bound for reasoning-chain compression. Building on an analysis of mainstream CoT compression strategies, we identify key structural factors enabling a model to approach MSL. Based on these insights, we propose TRiMS which employs the GRPO algorithm in conjunction with MSL-based estimation during training, while mitigating instabilities during the training process through dynamic batch aggregation and advantage computation using batch-level standard deviation. TRiMS achieves over 80% CoT token reduction with a minor accuracy boost across all benchmarks.

When a multi-agent system produces an incorrect or harmful answer, who is accountable if execution logs and agent identifiers are unavailable? Multi-agent language systems increasingly rely on structured interactions such as delegation and iterative refinement, yet the final output often obscures the underlying interaction topology and agent contributions. We introduce IET (Implicit Execution Tracing), a metadata-independent framework that enables token-level attribution directly from generated text and a simple mechanism for interaction topology reconstruction. During generation, agent-specific keyed signals are embedded into the token distribution, transforming the text into a self-describing execution trace detectable only with a secret key. At detection time, a transition-aware scoring method identifies agent handover points and reconstructs the interaction graph. Experiments show that IET recovers agent segments and coordination structure with high accuracy while preserving generation quality, enabling privacy-preserving auditing for multi-agent language systems.

GUI grounding is a critical capability for vision-language models (VLMs) that enables automated interaction with graphical user interfaces by locating target elements from natural language instructions. However, grounding on GUI screenshots remains challenging due to high-resolution images, small UI elements, and ambiguous user instructions. In this work, we propose AdaZoom-GUI, an adaptive zoom-based GUI grounding framework that improves both localization accuracy and instruction understanding. Our approach introduces an instruction refinement module that rewrites natural language commands into explicit and detailed descriptions, allowing the grounding model to focus on precise element localization. In addition, we design a conditional zoom-in strategy that selectively performs a second-stage inference on predicted small elements, improving localization accuracy while avoiding unnecessary computation and context loss on simpler cases. To support this framework, we construct a high-quality GUI grounding dataset and train the grounding model using Group Relative Policy Optimization (GRPO), enabling the model to predict both click coordinates and element bounding boxes. Experiments on public benchmarks demonstrate that our method achieves state-of-the-art performance among models with comparable or even larger parameter sizes, highlighting its effectiveness for high-resolution GUI understanding and practical GUI agent deployment.

Transformers enable in-context learning (ICL) for rapid, gradient-free adaptation in time series forecasting, yet most ICL-style approaches rely on tabularized, hand-crafted features, while end-to-end sequence models lack inference-time adaptation. We bridge this gap with a unified framework, Baguan-TS, which integrates the raw-sequence representation learning with ICL, instantiated by a 3D Transformer that attends jointly over temporal, variable, and context axes. To make this high-capacity model practical, we tackle two key hurdles: (i) calibration and training stability, improved with a feature-agnostic, target-space retrieval-based local calibration; and (ii) output oversmoothing, mitigated via context-overfitting strategy. On public benchmark with covariates, Baguan-TS consistently outperforms established baselines, achieving the highest win rate and significant reductions in both point and probabilistic forecasting metrics. Further evaluations across diverse real-world energy datasets demonstrate its robustness, yielding substantial improvements.

Non-stationarity is a fundamental challenge in multivariate long-term time series forecasting, often manifested as rapid changes in amplitude and phase. These variations lead to severe distribution shifts and consequently degrade predictive performance. Existing normalization-based methods primarily rely on first- and second-order statistics, implicitly assuming that distributions evolve smoothly and overlooking fine-grained temporal dynamics. To address these limitations, we propose TimeAPN, an Adaptive Amplitude-Phase Non-Stationarity Normalization framework that explicitly models and predicts non-stationary factors from both the time and frequency domains. Specifically, TimeAPN first models the mean sequence jointly in the time and frequency domains, and then forecasts its evolution over future horizons. Meanwhile, phase information is extracted in the frequency domain, and the phase discrepancy between the predicted and ground-truth future sequences is explicitly modeled to capture temporal misalignment. Furthermore, TimeAPN incorporates amplitude information into an adaptive normalization mechanism, enabling the model to effectively account for abrupt fluctuations in signal energy. The predicted non-stationary factors are subsequently integrated with the backbone forecasting outputs through a collaborative de-normalization process to reconstruct the final non-stationary time series. The proposed framework is model-agnostic and can be seamlessly integrated with various forecasting backbones. Extensive experiments on seven real-world multivariate datasets demonstrate that TimeAPN consistently improves long-term forecasting accuracy across multiple prediction horizons and outperforms state-of-the-art reversible normalization methods.

Lossless model compression holds tremendous promise for alleviating the memory and bandwidth bottlenecks in bit-exact Large Language Model (LLM) serving. However, existing approaches often result in substantial inference slowdowns due to fundamental design mismatches with GPU architectures: at the kernel level, variable-length bitstreams produced by traditional entropy codecs break SIMT parallelism; at the system level, decoupled pipelines lead to redundant memory traffic. We present ZipServ, a lossless compression framework co-designed for efficient LLM inference. ZipServ introduces Tensor-Core-Aware Triple Bitmap Encoding (TCA-TBE), a novel fixed-length format that enables constant-time, parallel decoding, together with a fused decompression-GEMM (ZipGEMM) kernel that decompresses weights on-the-fly directly into Tensor Core registers. This "load-compressed, compute-decompressed" design eliminates intermediate buffers and maximizes compute intensity. Experiments show that ZipServ reduces the model size by up to 30%, achieves up to 2.21x kernel-level speedup over NVIDIA's cuBLAS, and expedites end-to-end inference by an average of 1.22x over vLLM. ZipServ is the first lossless compression system that provides both storage savings and substantial acceleration for LLM inference on GPUs.

Transformer models have redefined sequence learning, yet dot-product self-attention introduces a quadratic token-mixing bottleneck for long-context time-series. We introduce the \textbf{Phasor Transformer} block, a phase-native alternative representing sequence states on the unit-circle manifold $S^1$. Each block combines lightweight trainable phase-shifts with parameter-free Discrete Fourier Transform (DFT) token coupling, achieving global $\mathcal{O}(N\log N)$ mixing without explicit attention maps. Stacking these blocks defines the \textbf{Large Phasor Model (LPM)}. We validate LPM on autoregressive time-series prediction over synthetic multi-frequency benchmarks. Operating with a highly compact parameter budget, LPM learns stable global dynamics and achieves competitive forecasting behavior compared to conventional self-attention baselines. Our results establish an explicit efficiency-performance frontier, demonstrating that large-model scaling for time-series can emerge from geometry-constrained phase computation with deterministic global coupling, offering a practical path toward scalable temporal modeling in oscillatory domains.

To think critically about arguments, human learners are trained to identify, reconstruct, and evaluate arguments. Argument reconstruction is especially important because it makes an argument's underlying inferences explicit. However, it remains unclear whether LLMs can similarly enhance their critical thinking ability by learning to reconstruct arguments. To address this question, we introduce a holistic framework with three contributions. We (1) propose an engine that automatically reconstructs arbitrary arguments (GAAR), (2) synthesize a new high-quality argument reconstruction dataset (Arguinas) using the GAAR engine, and (3) investigate whether learning argument reconstruction benefits downstream critical thinking tasks. Our experimental results show that, across seven critical thinking tasks, models trained to learn argument reconstruction outperform models that do not, with the largest performance gains observed when training on the proposed Arguinas dataset. The source code and dataset will be publicly available.

Most automated electronic medical record (EMR) pipelines remain output-oriented: they transcribe, extract, and summarize after the consultation, but they do not explicitly model what is already known, what is still missing, which uncertainty matters most, or what question or recommendation should come next. We formulate doctor-patient dialogue as a proactive knowledge-inquiry problem under partial observability. The proposed framework combines stateful extraction, sequential belief updating, gap-aware state modeling, hybrid retrieval over objectified medical knowledge, and a POMDP-lite action planner. Instead of treating the EMR as the only target artifact, the framework treats documentation as the structured projection of an ongoing inquiry loop. To make the formulation concrete, we report a controlled pilot evaluation on ten standardized multi-turn dialogues together with a 300-query retrieval benchmark aggregated across dialogues. On this pilot protocol, the full framework reaches 83.3% coverage, 80.0% risk recall, 81.4% structural completeness, and lower redundancy than the chunk-only and template-heavy interactive baselines. These pilot results do not establish clinical generalization; rather, they suggest that proactive inquiry may be methodologically interesting under tightly controlled conditions and can be viewed as a conceptually appealing formulation worth further investigation for dialogue-based EMR generation. This work should be read as a pilot concept demonstration under a controlled simulated setting rather than as evidence of clinical deployment readiness. No implication of clinical deployment readiness, clinical safety, or real-world clinical utility should be inferred from this pilot protocol.

Piecewise-linear nonlinear systems appear in many engineering disciplines. Prediction of the dynamic behavior of such systems is of great importance from practical and theoretical viewpoint. In this paper, a data-driven model order reduction method for piecewise-linear systems is proposed, which is based on dynamic mode decomposition (DMD). The overview of the concept of DMD is provided, and its application to model order reduction for nonlinear systems based on Galerkin projection is explained. The proposed approach uses impulse responses of the system to obtain snapshots of the state variables. The snapshots are then used to extract the dynamic modes that are used to form the projection basis vectors. The dynamics described by the equations of motion of the original full-order system are then projected onto the subspace spanned by the basis vectors. This produces a system with much smaller number of degrees of freedom (DOFs). The proposed method is applied to two representative examples of piecewise linear systems: a cantilevered beam subjected to an elastic stop at its end, and a bonded plates assembly with partial debonding. The reduced order models (ROMs) of these systems are constructed by using the Galerkin projection of the equation of motion with DMD modes alone, or DMD modes with a set of classical constraint modes to be able to handle the contact nonlinearity efficiently. The obtained ROMs are used for the nonlinear forced response analysis of the systems under harmonic loading. It is shown that the ROMs constructed by the proposed method produce accurate forced response results.

The rapid evolution toward 6G and beyond communication systems is accelerating the convergence of digital twins and world models at the network edge. Traditional digital twins provide high-fidelity representations of physical systems and support monitoring, analysis, and offline optimization. However, in highly dynamic edge environments, they face limitations in autonomy, adaptability, and scalability. This paper presents a systematic survey of the transition from digital twins to world models and discusses its role in enabling edge general intelligence (EGI). First, the paper clarifies the conceptual differences between digital twins and world models and highlights the shift from physics-based, centralized, and system-centric replicas to data-driven, decentralized, and agent-centric internal models. This discussion helps readers gain a clear understanding of how this transition enables more adaptive, autonomous, and resource-efficient intelligence at the network edge. The paper reviews the design principles, architectures, and key components of world models, including perception, latent state representation, dynamics learning, imagination-based planning, and memory. In addition, it examines the integration of world models and digital twins in wireless EGI systems and surveys emerging applications in integrated sensing and communications, semantic communication, air-ground networks, and low-altitude wireless networks. Finally, this survey provides a systematic roadmap and practical insights for designing world-model-driven edge intelligence systems in wireless and edge computing environments. It also outlines key research challenges and future directions toward scalable, reliable, and interoperable world models for edge-native agentic AI.

Autonomous AI agents powered by large language models are being deployed in production with capabilities including shell execution, file system access, database queries, and multi-party communication. Recent red teaming research demonstrates that these agents exhibit critical vulnerabilities in realistic settings: unauthorized compliance with non-owner instructions, sensitive information disclosure, identity spoofing, cross-agent propagation of unsafe practices, and indirect prompt injection through external resources [7]. In healthcare environments processing Protected Health Information, every such vulnerability becomes a potential HIPAA violation. This paper presents a security architecture deployed for nine autonomous AI agents in production at a healthcare technology company. We develop a six-domain threat model for agentic AI in healthcare covering credential exposure, execution capability abuse, network egress exfiltration, prompt integrity failures, database access risks, and fleet configuration drift. We implement four-layer defense in depth: (1) kernel level workload isolation using gVisor on Kubernetes, (2) credential proxy sidecars preventing agent containers from accessing raw secrets, (3) network egress policies restricting each agent to allowlisted destinations, and (4) a prompt integrity framework with structured metadata envelopes and untrusted content labeling. We report results from 90 days of deployment including four HIGH severity findings discovered and remediated by an automated security audit agent, progressive fleet hardening across three VM image generations, and defense coverage mapped to all eleven attack patterns from recent literature. All configurations, audit tooling, and the prompt integrity framework are released as open source.

Image registration is an ill-posed dense vision task, where multiple solutions achieve similar loss values, motivating probabilistic inference. Variational inference has previously been employed to capture these distributions, however restrictive assumptions about the posterior form can lead to poor characterisation, overconfidence and low-quality samples. More flexible posteriors are typically bottlenecked by the complexity of high-dimensional covariance matrices required for dense 3D image registration. In this work, we present a memory and computationally efficient inference method, Structured SIR, that enables expressive, multi-modal, characterisation of uncertainty with high quality samples. We propose the use of a Sampled Importance Resampling (SIR) algorithm with a novel memory-efficient high-dimensional covariance parameterisation as the sum of a low-rank covariance and a sparse, spatially structured Cholesky precision factor. This structure enables capturing complex spatial correlations while remaining computationally tractable. We evaluate the efficacy of this approach in 3D dense image registration of brain MRI data, which is a very high-dimensional problem. We demonstrate that our proposed methods produces uncertainty estimates that are significantly better calibrated than those produced by variational methods, achieving equivalent or better accuracy. Crucially, we show that the model yields highly structured multi-modal posterior distributions, enable effective and efficient uncertainty quantification.

Zero-shot learning (ZSL) aims to recognize the unseen classes in the open-world guided by the side-information (e.g., attributes). Its key task is how to infer the latent semantic knowledge between visual and attribute features on seen classes, and thus conducting a desirable semantic knowledge transfer from seen classes to unseen ones. Prior works simply utilize unidirectional attention within a weakly-supervised manner to learn the spurious and limited latent semantic representations, which fail to effectively discover the intrinsic semantic knowledge (e.g., attribute semantic) between visual and attribute features. To solve the above challenges, we propose a mutually causal semantic distillation network (termed MSDN++) to distill the intrinsic and sufficient semantic representations for ZSL. MSDN++ consists of an attribute$\rightarrow$visual causal attention sub-net that learns attribute-based visual features, and a visual$\rightarrow$attribute causal attention sub-net that learns visual-based attribute features. The causal attentions encourages the two sub-nets to learn causal vision-attribute associations for representing reliable features with causal visual/attribute learning. With the guidance of semantic distillation loss, the two mutual attention sub-nets learn collaboratively and teach each other throughout the training process. Extensive experiments on three widely-used benchmark datasets (e.g., CUB, SUN, AWA2, and FLO) show that our MSDN++ yields significant improvements over the strong baselines, leading to new state-of-the-art performances.

Recent diffusion-based extreme image compression methods have demonstrated remarkable performance at ultra-low bitrates. However, most approaches require training separate diffusion models for each target bitrate, resulting in substantial computational overhead and hindering practical deployment. Meanwhile, recent studies have shown that joint super-resolution can serve as an effective approach for enhancing low-bitrate reconstruction. However, when moving toward ultra-low bitrate regimes, these methods struggle due to severe information loss, and their reliance on fixed super-resolution scales prevents flexible adaptation across diverse bitrates. To address these limitations, we propose ASSR-EIC, a novel image compression framework that leverages arbitrary-scale super-resolution (ASSR) to support variable-rate extreme image compression (EIC). An arbitrary-scale downsampling module is introduced at the encoder side to provide controllable rate reduction, while a diffusion-based, joint degradation-aware ASSR decoder enables rate-adaptive reconstruction within a single model. We exploit the compression- and rescaling-aware diffusion prior to guide the reconstruction, yielding high fidelity and high realism restoration across diverse compression and rescaling settings. Specifically, we design a global compression-rescaling adaptor that offers holistic guidance for rate adaptation, and a local compression-rescaling modulator that dynamically balances generative and fidelity-oriented behaviors to achieve fine-grained, bitrate-adaptive detail restoration. To further enhance reconstruction quality, we introduce a dual semantic-enhanced design. Extensive experiments demonstrate that ASSR-EIC delivers state-of-the-art performance in extreme image compression while simultaneously supporting flexible bitrate control and adaptive rate-dependent reconstruction.

Causal representation learning (CRL) models aim to transform high-dimensional data into a latent space, enabling interventions to generate counterfactual samples or modify existing data based on the causal relationships among latent variables. To facilitate the development and evaluation of these models, a variety of synthetic and real-world datasets have been proposed, each with distinct advantages and limitations. For practical applications, CRL models must perform robustly across multiple evaluation directions, including reconstruction, disentanglement, causal discovery, and counterfactual reasoning, using appropriate metrics for each direction. However, this multi-directional evaluation can complicate model comparison, as a model may excel in some direction while under-performing in others. Another significant challenge in this field is reproducibility: the source code corresponding to published results must be publicly available, and repeated runs should yield performance consistent with the original reports. In this study, we critically analyzed the synthetic and real-world datasets currently employed in the literature, highlighting their limitations and proposing a set of essential characteristics for suitable datasets in CRL model development. We also introduce a single aggregate metric that consolidates performance across all evaluation directions, providing a comprehensive score for each model. Finally, we reviewed existing implementations from the literature and assessed them in terms of reproducibility, identifying gaps and best practices in the field.

Earthquake hazard analysis and design of spatially distributed infrastructure, such as power grids and energy pipeline networks, require scenario-specific ground-motion time histories with realistic frequency content and spatiotemporal coherence. However, producing the large ensembles needed for uncertainty quantification with physics-based simulations is computationally intensive and impractical for engineering workflows. To address this challenge, we introduce Ground-Motion Flow (GMFlow), a physics-inspired latent operator flow matching framework that generates realistic, large-scale regional ground-motion time-histories conditioned on physical parameters. Validated on simulated earthquake scenarios in the San Francisco Bay Area, GMFlow generates spatially coherent ground motion across more than 9 million grid points in seconds, achieving a 10,000-fold speedup over the simulation workflow, which opens a path toward rapid and uncertainty-aware hazard assessment for distributed infrastructure. More broadly, GMFlow advances mesh-agnostic functional generative modeling and could potentially be extended to the synthesis of large-scale spatiotemporal physical fields in diverse scientific domains.

A coding agent can bootstrap itself. Starting from a 926-word specification and a first implementation produced by an existing agent (Claude Code), a newly generated agent re-implements the same specification correctly from scratch. This reproduces, in the domain of AI coding agents, the classical bootstrap sequence known from compiler construction, and instantiates the meta-circular property known from Lisp. The result carries a practical implication: the specification, not the implementation, is the stable artifact of record. Improving an agent means improving its specification; the implementation is, in principle, regenerable at any time.

Free energies are fundamental quantities governing phase behavior and thermodynamic stability in polymer systems, yet their accurate computation often requires extensive simulations and post-processing techniques such as the Bennett Acceptance Ratio (BAR). While BAR provides reliable estimates when applied between closely related thermodynamic states, evaluating free energies across large changes in interaction strength typically requires a sequence of intermediate simulations to maintain sufficient phase-space overlap, substantially increasing computational cost. In this work we develop a machine learning framework for rapidly predicting excess free energies of linear diblock copolymer systems from simulation-derived energetic descriptors. Using dissipative particle dynamics simulations of freely-jointed chain polymers, we construct a dataset of per-chain energetic statistics, including heterogeneous interaction energies, homogeneous interaction energies, and bonded spring energies, and train feed-forward neural networks to learn the relationship between these descriptors and free energies computed using a stratified BAR procedure. The resulting models accurately reproduce the reference free energies across a range of chain lengths, compositions, and densities, including polymer architectures held out from training. In regimes where direct, brute-force BAR estimates become unreliable due to poor phase-space overlap, the neural network predictions remain consistent with the reference values. These results demonstrate that physically informed machine learning models can serve as efficient surrogates for expensive free-energy calculations and provide a promising approach for accelerating thermodynamic analysis of polymer systems.

The central challenge of reasoning-intensive retrieval lies in identifying implicitreasoning relationships between queries and documents, rather than superficial se-mantic or lexical similarity. The contrastive learning paradigm is fundamentallya static representation consolidation technique: during training, it encodes hier-archical relevance concepts into fixed geometric structures in the vector space,and at inference time it cannot dynamically adjust relevance judgments accord-ing to the specific reasoning demands of each query. Consequently, performancedegrades noticeably when vocabulary mismatch exists between queries and doc-uments or when implicit reasoning is required to establish relevance. This pa-per proposes Thought 1 (T1), a generative retrieval model that shifts relevancemodeling from static alignment to dynamic reasoning. On the query side, T1 dy-namically generates intermediate reasoning trajectories for each query to bridgeimplicit reasoning relationships and uses <embtoken> as a semantic aggregationpoint for the reasoning output. On the document side, it employs an instruction+ text + <embtoken> encoding format to support high-throughput indexing. Tointernalize dynamic reasoning capabilities into vector representations, we adopt athree-stage training curriculum and introduce GRPO in the third stage, enablingthe model to learn optimal derivation strategies for different queries through trial-and-error reinforcement learning. On the BRIGHT benchmark, T1-4B exhibitsstrong performance under the original query setting, outperforming larger modelstrained with contrastive learning overall, and achieving performance comparableto multi-stage retrieval pipelines. The results demonstrate that replacing static rep-resentation alignment with dynamic reasoning generation can effectively improvereasoning-intensive retrieval performance.

As retrieval models converge on generic benchmarks, the pressing question is no longer "who scores higher" but rather "where do systems fail, and why?" Person-job matching is a domain that urgently demands such diagnostic capability -- it requires systems not only to verify explicit constraints but also to perform skill-transfer inference and job-competency reasoning, yet existing benchmarks provide no systematic diagnostic support for this task. We introduce PJB (Person-Job Benchmark), a reasoning-aware retrieval evaluation dataset that uses complete job descriptions as queries and complete resumes as documents, defines relevance through job-competency judgment, is grounded in real-world recruitment data spanning six industry domains and nearly 200,000 resumes, and upgrades evaluation from "who scores higher" to "where do systems differ, and why" through domain-family and reasoning-type diagnostic labels. Diagnostic experiments using dense retrieval reveal that performance heterogeneity across industry domains far exceeds the gains from module upgrades for the same model, indicating that aggregate scores alone can severely mislead optimization decisions. At the module level, reranking yields stable improvements while query understanding not only fails to help but actually degrades overall performance when combined with reranking -- the two modules face fundamentally different improvement bottlenecks. The value of PJB lies not in yet another leaderboard of average scores, but in providing recruitment retrieval systems with a capability map that pinpoints where to invest.

Judea Pearl's do-calculus provides a foundation for causal inference, but its translation to continuous generative models remains fraught with geometric challenges. We establish the fundamental limits of such interventions. We define the Counterfactual Event Horizon and prove the Manifold Tearing Theorem: deterministic flows inevitably develop finite-time singularities under extreme interventions. We establish the Causal Uncertainty Principle for the trade-off between intervention extremity and identity preservation. Finally, we introduce Geometry-Aware Causal Flow (GACF), a scalable algorithm that utilizes a topological radar to bypass manifold tearing, validated on high-dimensional scRNA-seq data.

Current continuous generative models (e.g., Diffusion Models, Flow Matching) implicitly assume that locally consistent causal mechanisms naturally yield globally coherent counterfactuals. In this paper, we prove that this assumption fails fundamentally when the causal graph exhibits non-trivial homology (e.g., structural conflicts or hidden confounders). We formalize structural causal models as cellular sheaves over Wasserstein spaces, providing a strict algebraic topological definition of cohomological obstructions in measure spaces. To ensure computational tractability and avoid deterministic singularities (which we define as manifold tearing), we introduce entropic regularization and derive the Entropic Wasserstein Causal Sheaf Laplacian, a novel system of coupled non-linear Fokker-Planck equations. Crucially, we prove an entropic pullback lemma for the first variation of pushforward measures. By integrating this with the Implicit Function Theorem (IFT) on Sinkhorn optimality conditions, we establish a direct algorithmic bridge to automatic differentiation (VJP), achieving O(1)-memory reverse-mode gradients strictly independent of the iteration horizon. Empirically, our framework successfully leverages thermodynamic noise to navigate topological barriers ("entropic tunneling") in high-dimensional scRNA-seq counterfactuals. Finally, we invert this theoretical framework to introduce the Topological Causal Score, demonstrating that our Sheaf Laplacian acts as a highly sensitive algebraic detector for topology-aware causal discovery.

Virtual cell models aim to enable in silico experimentation by predicting how cells respond to genetic, chemical, or cytokine perturbations from single-cell measurements. In practice, however, large-scale perturbation prediction remains constrained by three coupled bottlenecks: inefficient training and inference pipelines, unstable modeling in high-dimensional sparse expression space, and evaluation protocols that overemphasize reconstruction-like accuracy while underestimating biological fidelity. In this work we present a specialized large-scale foundation model SCALE for virtual cell perturbation prediction that addresses the above limitations jointly. First, we build a BioNeMo-based training and inference framework that substantially improves data throughput, distributed scalability, and deployment efficiency, yielding 12.51* speedup on pretrain and 1.29* on inference over the prior SOTA pipeline under matched system settings. Second, we formulate perturbation prediction as conditional transport and implement it with a set-aware flow architecture that couples LLaMA-based cellular encoding with endpoint-oriented supervision. This design yields more stable training and stronger recovery of perturbation effects. Third, we evaluate the model on Tahoe-100M using a rigorous cell-level protocol centered on biologically meaningful metrics rather than reconstruction alone. On this benchmark, our model improves PDCorr by 12.02% and DE Overlap by 10.66% over STATE. Together, these results suggest that advancing virtual cells requires not only better generative objectives, but also the co-design of scalable infrastructure, stable transport modeling, and biologically faithful evaluation.

We develop an online learning algorithm that dramatically improves the data efficiency of reinforcement learning from human feedback (RLHF). Our algorithm incrementally updates reward and language models as choice data is received. The reward model is fit to the choice data, while the language model is updated by a variation of reinforce, with reinforcement signals provided by the reward model. Several features enable the efficiency gains: a small affirmative nudge added to each reinforcement signal, an epistemic neural network that models reward uncertainty, and information-directed exploration. With Gemma large language models (LLMs), our algorithm matches the performance of offline RLHF trained on 200K labels using fewer than 20K labels, representing more than a 10x gain in data efficiency. Extrapolating from our results, we expect our algorithm trained on 1M labels to match offline RLHF trained on 1B labels. This represents a 1,000x gain. To our knowledge, these are the first results to demonstrate that such large improvements are possible.

Large language models are rapidly being deployed as AI tutors, yet current evaluation paradigms assess problem-solving accuracy and generic safety in isolation, failing to capture whether a model is simultaneously pedagogically effective and safe across student-tutor interaction. We argue that tutoring safety is fundamentally different from conventional LLM safety: the primary risk is not toxic content but the quiet erosion of learning through answer over-disclosure, misconception reinforcement, and the abdication of scaffolding. To systematically study this failure mode, we introduce SafeTutors, a benchmark that jointly evaluates safety and pedagogy across mathematics, physics, and chemistry. SafeTutors is organized around a theoretically grounded risk taxonomy comprising 11 harm dimensions and 48 sub-risks drawn from learning-science literature. We uncover that all models show broad harm; scale doesn't reliably help; and multi-turn dialogue worsens behavior, with pedagogical failures rising from 17.7% to 77.8%. Harms also vary by subject, so mitigations must be discipline-aware, and single-turn "safe/helpful" results can mask systematic tutor failure over extended interaction.

Large vision-language models (VLMs) often exhibit weakened safety alignment with the integration of the visual modality. Even when text prompts contain explicit harmful intent, adding an image can substantially increase jailbreak success rates. In this paper, we observe that VLMs can clearly distinguish benign inputs from harmful ones in their representation space. Moreover, even among harmful inputs, jailbreak samples form a distinct internal state that is separable from refusal samples. These observations suggest that jailbreaks do not arise from a failure to recognize harmful intent. Instead, the visual modality shifts representations toward a specific jailbreak state, thereby leading to a failure to trigger refusal. To quantify this transition, we identify a jailbreak direction and define the jailbreak-related shift as the component of the image-induced representation shift along this direction. Our analysis shows that the jailbreak-related shift reliably characterizes jailbreak behavior, providing a unified explanation for diverse jailbreak scenarios. Finally, we propose a defense method that enhances VLM safety by removing the jailbreak-related shift (JRS-Rem) at inference time. Experiments show that JRS-Rem provides strong defense across multiple scenarios while preserving performance on benign tasks.

Large reasoning models (LRMs) achieved remarkable performance via chain-of-thought (CoT), but recent studies showed that such enhanced reasoning capabilities are at the expense of significantly degraded safety capabilities. In this paper, we reveal that LRMs' safety degradation occurs only after CoT is enabled, and this degradation is not observed when CoT is disabled. This observation motivates us to consider encouraging LRMs to make safety decisions before CoT generation. To this end, we propose a novel safety alignment method that promotes the safety decision-making of LRMs before starting CoT generation. Specifically, we first utilize a Bert-based classifier to extract safety decision signals from a safe model (e.g., a CoT-disabled LRM) and then integrate these signals into LRMs' safety alignment as auxiliary supervision. In this way, the safety gradients can be backpropagated to the LRMs' latent representations, effectively strengthening the LRMs' safety decision-making abilities against CoT generation. Extensive experiments demonstrate that our method substantially improves the safety capabilities of LRMs while effectively maintaining LRMs' general reasoning performance.

Internal noise in deep networks is usually inherited from heuristics such as dropout, hard masking, or additive perturbation. We ask two questions: what correlation geometry should internal noise have, and is the implemented perturbation compatible with the representations it acts on? We answer these questions through Variational Kernel Design (VKD), a framework in which a noise mechanism is specified by a law family, a correlation kernel, and an injection operator, and is derived from learning desiderata. In a solved spatial subfamily, a quadratic maximum-entropy principle over latent log-fields yields a Gaussian optimizer with precision given by the Dirichlet Laplacian, so the induced geometry is the Dirichlet Green kernel. Wick normalization then gives a canonical positive mean-one gate, Gaussian Chaos Noise (GCh). For the sample-wise gate used in practice, we prove exact Gaussian control of pairwise log-ratio deformation, margin-sensitive ranking stability, and an exact expected intrinsic roughness budget; hard binary masks instead induce singular or coherence-amplified distortions on positive coherent representations. On ImageNet and ImageNet-C, GCh consistently improves calibration and under shift also improves NLL at competitive accuracy.

Proper citation of relevant literature is essential for contextualising and validating scientific contributions. While current citation recommendation systems leverage local and global textual information, they often overlook the nuances of the human citation behaviour. Recent methods that incorporate such patterns improve performance but incur high computational costs and introduce systematic biases into downstream rerankers. To address this, we propose Profiler, a lightweight, non-learnable module that captures human citation patterns efficiently and without bias, significantly enhancing candidate retrieval. Furthermore, we identify a critical limitation in current evaluation protocol: the systems are assessed in a transductive setting, which fails to reflect real-world scenarios. We introduce a rigorous Inductive evaluation setting that enforces strict temporal constraints, simulating the recommendation of citations for newly authored papers in the wild. Finally, we present DAVINCI, a novel reranking model that integrates profiler-derived confidence priors with semantic information via an adaptive vector-gating mechanism. Our system achieves new state-of-the-art results across multiple benchmark datasets, demonstrating superior efficiency and generalisability.

Computer use agents create new privacy risks: training data collected from real websites inevitably contains sensitive information, and cloud-hosted inference exposes user screenshots. Detecting personally identifiable information in web screenshots is critical for privacy-preserving deployment, but no public benchmark exists for this task. We introduce WebPII, a fine-grained synthetic benchmark of 44,865 annotated e-commerce UI images designed with three key properties: extended PII taxonomy including transaction-level identifiers that enable reidentification, anticipatory detection for partially-filled forms where users are actively entering data, and scalable generation through VLM-based UI reproduction. Experiments validate that these design choices improve layout-invariant detection across diverse interfaces and generalization to held-out page types. We train WebRedact to demonstrate practical utility, more than doubling text-extraction baseline accuracy (0.753 vs 0.357 mAP@50) at real-time CPU latency (20ms). We release the dataset and model to support privacy-preserving computer use research.

Drug recommendation requires a deep understanding of individual patient context, especially for complex conditions like Parkinson's disease. While LLMs possess broad medical knowledge, they fail to capture the subtle nuances of actual prescribing patterns. Existing RAG methods also struggle with these complexities because guideline-based retrieval remains too generic and similar-patient retrieval often replicates majority patterns without accounting for the unique clinical nuances of individual patients. To bridge this gap, we propose PACE-RAG (Patient-Aware Contextual and Evidence-based Policy RAG), a novel framework designed to synthesize individual patient context with the prescribing tendencies of similar cases. By analyzing treatment patterns tailored to specific clinical signals, PACE-RAG identifies optimal prescriptions and generates an explainable clinical summary. Evaluated on a Parkinson's cohort and the MIMIC-IV benchmark using Llama-3.1-8B and Qwen3-8B, PACE-RAG achieved state-of-the-art performance, reaching F1 scores of 80.84% and 47.22%, respectively. These results validate PACE-RAG as a robust, clinically grounded solution for personalized decision support. Our code is available at: https://github.com/ChaeYoungHuh/PACE-RAG.

Layer-wise mixed-precision quantization (LMPQ) enables effective compression under extreme low-bit settings by allocating higher precision to sensitive layers. However, existing methods typically treat all intra-layer weight modules uniformly and rely on a single numerical property when estimating sensitivity, overlooking their distinct operational roles and structural characteristics. To address this, we propose NSDS, a novel calibration-free LMPQ framework driven by Numerical and Structural Dual-Sensitivity. Specifically, it first mechanistically decomposes each layer into distinct operational roles and quantifies their sensitivity from both numerical and structural perspectives. These dual-aspect scores are then aggregated into a unified layer-wise metric through a robust aggregation scheme based on MAD-Sigmoid and Soft-OR to guide bit allocation. Extensive experiments demonstrate that NSDS consistently achieves superior performance compared to various baselines across diverse models and downstream tasks, without relying on any calibration data.

The finite symmetric group S_n provides a natural domain for permutations, yet learning probability distributions on S_n is challenging due to its factorially growing size and discrete, non-Euclidean structure. Recent permutation diffusion methods define forward noising via shuffle-based random walks (e.g., riffle shuffles) and learn reverse transitions with Plackett-Luce (PL) variants, but the resulting trajectories can be abrupt and increasingly hard to denoise as n grows. We propose Soft-Rank Diffusion, a discrete diffusion framework that replaces shuffle-based corruption with a structured soft-rank forward process: we lift permutations to a continuous latent representation of order by relaxing discrete ranks into soft ranks, yielding smoother and more tractable trajectories. For the reverse process, we introduce contextualized generalized Plackett-Luce (cGPL) denoisers that generalize prior PL-style parameterizations and improve expressivity for sequential decision structures. Experiments on sorting and combinatorial optimization benchmarks show that Soft-Rank Diffusion consistently outperforms prior diffusion baselines, with particularly strong gains in long-sequence and intrinsically sequential settings.

We introduce GSU, a text-only grid dataset to evaluate the spatial reasoning capabilities of LLMs over 3 core tasks: navigation, object localization, and structure composition. By forgoing visual inputs, isolating spatial reasoning from perception, we show that while most models grasp basic grid concepts, they struggle with frames of reference relative to an embodied agent and identifying 3D shapes from coordinate lists. We also find that exposure to a visual modality does not provide a generalizable understanding of 3D space that VLMs are able to utilize for these tasks. Finally, we show that while the very latest frontier models can solve the provided tasks (though harder variants may still stump them), fully fine-tuning a small LM or LORA fine-tuning a small LLM show potential to match frontier model performance, suggesting an avenue for specialized embodied agents.

Machine Learning (ML) cloud services, offered by leading providers such as Amazon, Google, and Microsoft, enable the integration of ML components into software systems without building models from scratch. However, the rapid adoption of ML services, coupled with the growing complexity of business requirements, has led to widespread misuses, compromising the quality, maintainability, and evolution of ML service-based systems. Though prior research has studied patterns and antipatterns in service-based and ML-based systems separately, automatic detection of ML service misuses remains a challenge. In this paper, we propose MLmisFinder, an automatic approach to detect ML service misuses in software systems, aiming to identify instances of improper use of ML services to help developers properly integrate ML components in ML service-based systems. We propose a metamodel that captures the data needed to detect misuses in ML service-based systems and apply a set of rule-based detection algorithms for seven misuse types. We evaluated MLmisFinder on 107 software systems collected from open-source GitHub repositories and compared it with a state-of-the-art baseline. Our results show that MLmisFinder effectively detects ML service misuses, achieving an average precision of 96.7\% and recall of 97\%, outperforming the state-of-the-art baseline. MLmisFinder also scaled efficiently to detect misuses across 817 ML service-based systems and revealed that such misuses are widespread, especially in areas such as data drift monitoring and schema validation.

The efficient adjudication of responsibility disputes is pivotal for maintaining marketplace fairness. However, the exponential surge in ride-hailing volume renders manual review intractable, while conventional automated methods lack the reasoning transparency required for quasi-judicial decisions. Although Multimodal LLMs offer a promising paradigm, they fundamentally struggle to bridge the gap between general visual semantics and rigorous evidentiary protocols, often leading to perceptual hallucinations and logical looseness. To address these systemic misalignments, we introduce RideJudge, a Progressive Visual-Logic-Aligned Framework. Instead of relying on generic pre-training, we bridge the semantic gap via SynTraj, a synthesis engine that grounds abstract liability concepts into concrete trajectory patterns. To resolve the conflict between massive regulation volume and limited context windows, we propose an Adaptive Context Optimization strategy that distills expert knowledge, coupled with a Chain-of-Adjudication mechanism to enforce active evidentiary inquiry. Furthermore, addressing the inadequacy of sparse binary feedback for complex liability assessment, we implement a novel Ordinal-Sensitive Reinforcement Learning mechanism that calibrates decision boundaries against hierarchical severity. Extensive experiments show that our RideJudge-8B achieves 88.41\% accuracy, surpassing 32B-scale baselines and establishing a new standard for interpretable adjudication.

We present ShuttleEnv, an interactive and data-driven simulation environment for badminton, designed to support reinforcement learning and strategic behavior analysis in fast-paced adversarial sports. The environment is grounded in elite-player match data and employs explicit probabilistic models to simulate rally-level dynamics, enabling realistic and interpretable agent-opponent interactions without relying on physics-based simulation. In this demonstration, we showcase multiple trained agents within ShuttleEnv and provide live, step-by-step visualization of badminton rallies, allowing attendees to explore different play styles, observe emergent strategies, and interactively analyze decision-making behaviors. ShuttleEnv serves as a reusable platform for research, visualization, and demonstration of intelligent agents in sports AI. Our ShuttleEnv demo video URL: https://drive.google.com/file/d/1hTR4P16U27H2O0-w316bR73pxE2ucczX/view

International shipping produces approximately 3% of global greenhouse gas emissions, yet voyage routing remains dominated by heuristic methods. We present PIER (Physics-Informed, Energy-efficient, Risk-aware routing), an offline reinforcement learning framework that learns fuel-efficient, safety-aware routing policies from physics-calibrated environments grounded in historical vessel tracking data and ocean reanalysis products, requiring no online simulator. Validated on one full year (2023) of AIS data across seven Gulf of Mexico routes (840 episodes per method), PIER reduces mean CO2 emissions by 10% relative to great-circle routing. However, PIER's primary contribution is eliminating catastrophic fuel waste: great-circle routing incurs extreme fuel consumption (>1.5x median) in 4.8% of voyages; PIER reduces this to 0.5%, a 9-fold reduction. Per-voyage fuel variance is 3.5x lower (p<0.001), with bootstrap 95% CI for mean savings [2.9%, 15.7%]. Partial validation against observed AIS vessel behavior confirms consistency with the fastest real transits while exhibiting 23.1x lower variance. Crucially, PIER is forecast-independent: unlike A* path optimization whose wave protection degrades 4.5x under realistic forecast uncertainty, PIER maintains constant performance using only local observations. The framework combines physics-informed state construction, demonstration-augmented offline data, and a decoupled post-hoc safety shield, an architecture that transfers to wildfire evacuation, aircraft trajectory optimization, and autonomous navigation in unmapped terrain.

Accurately estimating task progress is critical for embodied agents to plan and execute long-horizon, multi-step tasks. Despite promising advances, existing Vision-Language Models (VLMs) based methods primarily leverage their video understanding capabilities, while neglecting their complex reasoning potential. Furthermore, processing long video trajectories with VLMs is computationally prohibitive for real-world deployment. To address these challenges, we propose the Recurrent Reasoning Vision-Language Model ($\text{R}^2$VLM). Our model features a recurrent reasoning framework that processes local video snippets iteratively, maintaining a global context through an evolving Chain of Thought (CoT). This CoT explicitly records task decomposition, key steps, and their completion status, enabling the model to reason about complex temporal dependencies. This design avoids the high cost of processing long videos while preserving essential reasoning capabilities. We train $\text{R}^2$VLM on large-scale, automatically generated datasets from ALFRED and Ego4D. Extensive experiments on progress estimation and downstream applications, including progress-enhanced policy learning, reward modeling for reinforcement learning, and proactive assistance, demonstrate that $\text{R}^2$VLM achieves strong performance and generalization, achieving a new state-of-the-art in long-horizon task progress estimation. The models and benchmarks are publicly available at \href{https://huggingface.co/collections/zhangyuelin/r2vlm}{huggingface}.

We present Ruyi2.5, a multimodal familial model built on the AI Flow framework. Extending Ruyi2's "Train Once, Deploy Many" paradigm to the multimodal domain, Ruyi2.5 constructs a shared-backbone architecture that co-trains models of varying scales within a single unified pipeline, ensuring semantic consistency across all deployment tiers. Built upon Ruyi2.5, Ruyi2.5-Camera model is developed as a privacy-preserving camera service system, which instantiates Ruyi2.5-Camera into a two-stage recognition pipeline: an edge model applies information-bottleneck-guided irreversible feature mapping to de-identify raw frames at the source, while a cloud model performs deep behavior reasoning. To accelerate reinforcement learning fine-tuning, we further propose Binary Prefix Policy Optimization (BPPO), which reduces sample redundancy via binary response selection and focuses gradient updates on response prefixes, achieving a 2 to 3 times training speedup over GRPO. Experiments show Ruyi2.5 matches Qwen3-VL on the general multimodal benchmarks, while Ruyi2.5-Camera substantially outperforms Qwen3-VL on privacy-constrained surveillance tasks.

Large Language Models (LLMs) with extended reasoning capabilities often generate verbose and redundant reasoning traces, incurring unnecessary computational cost. While existing reinforcement learning approaches address this by optimizing final response length, they neglect the quality of intermediate reasoning steps, leaving models vulnerable to reward hacking. We argue that verbosity is not merely a length problem, but a symptom of poor intermediate reasoning quality. To investigate this, we conduct an empirical study tracking the conditional entropy of the answer distribution across reasoning steps. We find that high-quality reasoning traces exhibit two consistent properties: low uncertainty convergence and monotonic progress. These findings suggest that high-quality reasoning traces are informationally dense, that is, each step contributes meaningful entropy reduction relative to the total reasoning length. Motivated by this, we propose InfoDensity, a reward framework for RL training that combines an AUC-based reward and a monotonicity reward as a unified measure of reasoning quality, weighted by a length scaling term that favors achieving equivalent quality more concisely. Experiments on mathematical reasoning benchmarks demonstrate that InfoDensity matches or surpasses state-of-the-art baselines in accuracy while significantly reducing token usage, achieving a strong accuracy-efficiency trade-off.

Reducing latency and energy consumption is critical to improving the efficiency of memory systems in modern computing. This work introduces ReLMXEL (Reinforcement Learning for Memory Controller with Explainable Energy and Latency Optimization), a explainable multi-agent online reinforcement learning framework that dynamically optimizes memory controller parameters using reward decomposition. ReLMXEL operates within the memory controller, leveraging detailed memory behavior metrics to guide decision-making. Experimental evaluations across diverse workloads demonstrate consistent performance gains over baseline configurations, with refinements driven by workload-specific memory access behaviour. By incorporating explainability into the learning process, ReLMXEL not only enhances performance but also increases the transparency of control decisions, paving the way for more accountable and adaptive memory system designs.

Despite rapid developments and widespread applications of MLLM agents, they still struggle with long-form video understanding (LVU) tasks, which are characterized by high information density and extended temporal spans. Recent research on LVU agents demonstrates that simple task decomposition and collaboration mechanisms are insufficient for long-chain reasoning tasks. Moreover, directly reducing the time context through embedding-based retrieval may lose key information of complex problems. In this paper, we propose Symphony, a multi-agent system, to alleviate these limitations. By emulating human cognition patterns, Symphony decomposes LVU into fine-grained subtasks and incorporates a deep reasoning collaboration mechanism enhanced by reflection, effectively improving the reasoning capability. Additionally, Symphony provides a VLM-based grounding approach to analyze LVU tasks and assess the relevance of video segments, which significantly enhances the ability to locate complex problems with implicit intentions and large temporal spans. Experimental results show that Symphony achieves state-of-the-art performance on LVBench, LongVideoBench, VideoMME, and MLVU, with a 5.0% improvement over the prior state-of-the-art method on LVBench. Code is available at https://github.com/Haiyang0226/Symphony.

A foundational assumption in linguistics holds that the relationship between a word's sound and its meaning is arbitrary. Accumulating evidence from sound symbolism challenges this view, yet no study has systematically mapped the multidimensional semantic profile of every phonological unit within a language. Here we show that individual letter-phonemes in English carry structured, multidimensional semantic signals. Using a minimal-pair paradigm spanning all 220 pairwise letter contrasts, three large language models independently recover consistent phoneme-meaning associations across nine perceptual dimensions. These associations are systematically predicted by articulatory-phonetic features, with manner and place of articulation mapping onto distinct semantic dimensions. Behavioral data from English speakers confirm these patterns at rates well above chance (80.8%), and preliminary cross-linguistic evidence from five typologically diverse languages suggests that core mappings generalize beyond English. Our findings indicate that sound-meaning iconicity is not an occasional curiosity but a pervasive, structured property of the phonological signal, one so systematic that large language models recover it when given only text input, without exposure to speech or articulation during the task.

We propose CRAFT, a red-teaming alignment framework that leverages model reasoning capabilities and hidden representations to improve robustness against jailbreak attacks. Unlike prior defenses that operate primarily at the output level, CRAFT aligns large reasoning models to generate safety-aware reasoning traces by explicitly optimizing objectives defined over the hidden state space. Methodologically, CRAFT integrates contrastive representation learning with reinforcement learning to separate safe and unsafe reasoning trajectories, yielding a latent-space geometry that supports robust, reasoning-level safety alignment. Theoretically, we show that incorporating latent-textual consistency into GRPO eliminates superficially aligned policies by ruling them out as local optima. Empirically, we evaluate CRAFT on multiple safety benchmarks using two strong reasoning models, Qwen3-4B-Thinking and R1-Distill-Llama-8B, where it consistently outperforms state-of-the-art defenses such as IPO and SafeKey. Notably, CRAFT delivers an average 79.0% improvement in reasoning safety and 87.7% improvement in final-response safety over the base models, demonstrating the effectiveness of hidden-space reasoning alignment.

Culture-expressions, such as idioms, slang, and culture-specific items (CSIs), are pervasive in natural language and encode meanings that go beyond literal linguistic form. Accurately translating such expressions remains challenging for machine translation systems. Despite this, existing benchmarks remain fragmented and do not provide a systematic framework for evaluating translation performance on culture-loaded expressions. To address this gap, we introduce CulT-Eval, a benchmark designed to evaluate how models handle different types of culturally grounded expressions. CulT-Eval comprises over 7,959 carefully curated instances spanning multiple types of culturally grounded expressions, with a comprehensive error taxonomy covering culturally grounded expressions. Through extensive evaluation of large language models and detailed analysis, we identify recurring and systematic failure modes that are not adequately captured by existing automatic metrics. Accordingly, we propose a complementary evaluation metric that targets culturally induced meaning deviations overlooked by standard MT metrics. The results indicate that current models struggle to preserve culturally grounded meaning and to capture the cultural and contextual nuances essential for accurate translation. Our benchmark and code are available at https://anonymous.4open.science/r/CulT-Eval-E75D/.

Generative Flow Networks for continuous scenarios (CFlowNets) have shown promise in solving sequential decision-making tasks by learning stochastic policies using a flow and a retrieval network. Despite their demonstrated efficiency compared to state-of-the-art Reinforcement Learning (RL) algorithms, their practical application in robotic control tasks is constrained by the reliance on pre-training the retrieval network. This dependency poses challenges in dynamic robotic environments, where pre-training data may not be readily available or representative of the current environment. This paper introduces WINFlowNets, a novel CFlowNets framework that enables the co-training of flow and retrieval networks. WINFlowNets begins with a warm-up phase for the retrieval network to bootstrap its policy, followed by a shared training architecture and a shared replay buffer for co-training both networks. Experiments in simulated robotic environments demonstrate that WINFlowNets surpasses CFlowNets and state-of-the-art RL algorithms in terms of average reward and training stability. Furthermore, WINFlowNets exhibits strong adaptive capability in fault environments, making it suitable for tasks that demand quick adaptation with limited sample data. These findings highlight WINFlowNets' potential for deployment in dynamic and malfunction-prone robotic systems, where traditional pre-training or sample inefficient data collection may be impractical.

GenAI Units In Digital Design Education (GUIDE) is an open courseware repository with runnable Google Colab labs and other materials. We describe the repository's architecture and educational approach based on standardized teaching units comprising slides, short videos, runnable labs, and related papers. This organization enables consistency for both the students' learning experience and the reuse and grading by instructors. We demonstrate GUIDE in practice with three representative units: VeriThoughts for reasoning and formal-verification-backed RTL generation, enhanced LLM-aided testbench generation, and LLMPirate for IP Piracy. We also provide details for four example course instances (GUIDE4ChipDesign, Build your ASIC, GUIDE4HardwareSecurity, and Hardware Design) that assemble GUIDE units into full semester offerings, learning outcomes, and capstone projects, all based on proven materials. For example, the GUIDE4HardwareSecurity course includes a project on LLM-aided hardware Trojan insertion that has been successfully deployed in the classroom and in Cybersecurity Games and Conference (CSAW), a student competition and academic conference for cybersecurity. We also organized an NYU Cognichip Hackathon, engaging students across 24 international teams in AI-assisted RTL design workflows. The GUIDE repository is open for contributions and available at: https://github.com/FCHXWH823/LLM4ChipDesign.

Story visualization requires generating sequential imagery that aligns semantically with evolving narratives while maintaining rigorous consistency in character identity and visual style. However, existing methodologies often struggle with subject inconsistency and identity drift, particularly when depicting complex interactions or extended narrative arcs. To address these challenges, we propose a cohesive two-stage framework designed for robust and consistent story generation. First, we introduce Group-Shared Attention (GSA), a mechanism that fosters intrinsic consistency by enabling lossless cross-sample information flow within attention layers. This allows the model to structurally encode identity correspondence across frames without relying on external encoders. Second, we leverage Direct Preference Optimization (DPO) to align generated outputs with human aesthetic and narrative standards. Unlike conventional methods that rely on conflicting auxiliary losses, our approach simultaneously enhances visual fidelity and identity preservation by learning from holistic preference data. Extensive evaluations on the ViStoryBench benchmark demonstrate that our method establishes a new state-of-the-art, significantly outperforming strong baselines with gains of +10.0 in Character Identity (CIDS) and +18.7 in Style Consistency (CSD), all while preserving high-fidelity generation.

Ordinal data is widely prevalent in clinical and other domains, yet there is a lack of both modern, machine-learning based methods and publicly available software to address it. In this paper, we present a model-agnostic method of ordinal classification, which can apply any non-ordinal classification method in an ordinal fashion. We also provide an open-source implementation of these algorithms, in the form of a Python package. We apply these models on multiple real-world datasets to show their performance across domains. We show that they often outperform non-ordinal classification methods, especially when the number of datapoints is relatively small or when there are many classes of outcomes. This work, including the developed software, facilitates the use of modern, more powerful machine learning algorithms to handle ordinal data.

Modern convolutional neural networks (CNNs) are workhorses for video and image processing, but fail to adapt to the computational complexity of input samples in a dynamic manner to minimize energy consumption. In this research, we propose DANCE, a fine-grained, input-aware, dynamic pruning framework for 3D CNNs to maximize power efficiency with negligible to zero impact on performance. In the proposed two-step approach, the first step is called activation variability amplification (AVA), and the 3D CNN model is retrained to increase the variance of the magnitude of neuron activations across the network in this step, facilitating pruning decisions across diverse CNN input scenarios. In the second step, called adaptive activation pruning (AAP), a lightweight activation controller network is trained to dynamically prune frames, channels, and features of 3D convolutional layers of the network (different for each layer), based on statistics of the outputs of the first layer of the network. Our method achieves substantial savings in multiply-accumulate (MAC) operations and memory accesses by introducing sparsity within convolutional layers. Hardware validation on the NVIDIA Jetson Nano GPU and the Qualcomm Snapdragon 8 Gen 1 platform demonstrates respective speedups of 1.37X and 2.22X, achieving up to 1.47X higher energy efficiency compared to the state of the art.

Gaussian process (GP) regression is widely used for uncertainty quantification, yet the standard formulation assumes noise-free covariates. When inputs are measured with error, this errors-in-variables (EIV) setting can lead to optimistically narrow posterior intervals and biased decisions. We study GP regression under input measurement uncertainty by representing each noisy input as a probability measure and defining covariance through Wasserstein distances between these measures. Building on this perspective, we instantiate a deterministic projected Wasserstein ARD (PWA) kernel whose one-dimensional components admit closed-form expressions and whose product structure yields a scalable, positive-definite kernel on distributions. Unlike latent-input GP models, PWA-based GPs (\PWAGPs) handle input noise without introducing unobserved covariates or Monte Carlo projections, making uncertainty quantification more transparent and robust.

Attacks can exploit zero-day or one-day vulnerabilities that are not publicly disclosed. To detect these vulnerabilities, security researchers monitor development activities in open-source repositories to identify unreported security patches. The sheer volume of commits makes this task infeasible to accomplish manually. Consequently, security patch detectors commonly trained and evaluated on security patches linked from vulnerability reports in the National Vulnerability Database (NVD). In this study, we assess the effectiveness of these detectors when applied in-the-wild. Our results show that models trained on NVD-derived data show substantially decreased performance, with decreases in F1-score of up to 90\% when tested on in-the-wild security patches, rendering them impractical for real-world use. An analysis comparing security patches identified in-the-wild and commits linked from NVD reveals that they can be easily distinguished from each other. Security patches associated with NVD have different distribution of commit messages, vulnerability types, and composition of changes. These differences suggest that NVD may be unsuitable as the \textit{sole} source of data for training models to detect security patches. We find that constructing a dataset that combines security patches from NVD data with a small subset of manually identified security patches can improve model robustness.

Recent advances in Large Language Models (LLMs) and Large Multimodal Models (LMMs) have improved Document Layout Analysis (DLA), yet structural errors such as region merging, splitting, and omission remain persistent. Conventional overlap-based metrics (e.g., IoU, mAP) fail to capture such logical inconsistencies. To overcome this limitation, we propose Layout Error Detection (LED), a benchmark that evaluates structural reasoning in DLA predictions beyond surface-level accuracy. LED defines eight standardized error types (Missing, Hallucination, Size Error, Split, Merge, Overlap, Duplicate, and Misclassification) and provides quantitative rules and injection algorithms for realistic error simulation. Using these definitions, we construct LED-Dataset and design three evaluation tasks: document-level error detection, document-level error-type classification, and element-level error-type classification. Experiments with state-of-the-art multimodal models show that LED enables fine-grained and interpretable assessment of structural understanding, revealing clear weaknesses across modalities and architectures. Overall, LED establishes a unified and explainable benchmark for diagnosing the structural robustness and reasoning capability of document understanding models.

Label noise has been broadly observed in real-world datasets. To mitigate the negative impact of overfitting to label noise for deep models, effective strategies (\textit{e.g.}, re-weighting, or loss rectification) have been broadly applied in prevailing approaches, which have been generally learned under the meta-learning scenario. Despite the robustness of noise achieved by the probabilistic meta-learning models, they usually suffer from model collapse that degenerates generalization performance. In this paper, we propose variational rectification inference (VRI) to formulate the adaptive rectification for loss functions as an amortized variational inference problem and derive the evidence lower bound under the meta-learning framework. Specifically, VRI is constructed as a hierarchical Bayes by treating the rectifying vector as a latent variable, which can rectify the loss of the noisy sample with the extra randomness regularization and is, therefore, more robust to label noise. To achieve the inference of the rectifying vector, we approximate its conditional posterior with an amortization meta-network. By introducing the variational term in VRI, the conditional posterior is estimated accurately and avoids collapsing to a Dirac delta function, which can significantly improve the generalization performance. The elaborated meta-network and prior network adhere to the smoothness assumption, enabling the generation of reliable rectification vectors. Given a set of clean meta-data, VRI can be efficiently meta-learned within the bi-level optimization programming. Besides, theoretical analysis guarantees that the meta-network can be efficiently learned with our algorithm. Comprehensive comparison experiments and analyses validate its effectiveness for robust learning with noisy labels, particularly in the presence of open-set noise.

Reconstructing a 12-lead electrocardiogram (ECG) from a reduced lead set is an ill-posed inverse problem due to anatomical variability. Standard deep learning methods often ignore underlying cardiac pathology losing vital morphology in precordial leads. We propose Pathology-Aware Multi-View Contrastive Learning, a framework that regularizes the latent space through a pathological manifold. Our architecture integrates high-fidelity time-domain waveforms with pathology-aware embeddings learned via supervised contrastive alignment. By maximizing mutual information between latent representations and clinical labels, the framework learns to filter anatomical "nuisance" variables. On the PTB-XL dataset, our method achieves approx. 76\% reduction in RMSE compared to state-of-the-art model in patient-independent setting. Cross-dataset evaluation on the PTB Diagnostic Database confirms superior generalization, bridging the gap between hardware portability and diagnostic-grade reconstruction.

We propose Q-BIOLAT, a framework for modeling and optimizing protein fitness landscapes in binary latent spaces. Starting from protein sequences, we leverage pretrained protein language models to obtain continuous embeddings, which are then transformed into compact binary latent representations. In this space, protein fitness is approximated using a quadratic unconstrained binary optimization (QUBO) model, enabling efficient combinatorial search via classical heuristics such as simulated annealing and genetic algorithms. On the ProteinGym benchmark, we demonstrate that Q-BIOLAT captures meaningful structure in protein fitness landscapes and enables the identification of high-fitness variants. Despite using a simple binarization scheme, our method consistently retrieves sequences whose nearest neighbors lie within the top fraction of the training fitness distribution, particularly under the strongest configurations. We further show that different optimization strategies exhibit distinct behaviors, with evolutionary search performing better in higher-dimensional latent spaces and local search remaining competitive in preserving realistic sequences. Beyond its empirical performance, Q-BIOLAT provides a natural bridge between protein representation learning and combinatorial optimization. By formulating protein fitness as a QUBO problem, our framework is directly compatible with emerging quantum annealing hardware, opening new directions for quantum-assisted protein engineering. Our implementation is publicly available at: https://github.com/HySonLab/Q-BIOLAT

Vision-Language Models (VLMs) exhibit a characteristic "cone effect" in which nonlinear encoders map embeddings into highly concentrated regions of the representation space, contributing to cross-modal separation known as the modality gap. While this phenomenon has been widely observed, its practical impact on supervised multimodal learning -particularly in medical domains- remains unclear. In this work, we introduce a lightweight post-hoc mechanism that keeps pretrained VLM encoders frozen while continuously controlling cross-modal separation through a single hyperparameter {λ}. This enables systematic analysis of how the modality gap affects downstream multimodal performance without expensive retraining. We evaluate generalist (CLIP, SigLIP) and medically specialized (BioMedCLIP, MedSigLIP) models across diverse medical and natural datasets in a supervised multimodal settings. Results consistently show that reducing excessive modality gap improves downstream performance, with medical datasets exhibiting stronger sensitivity to gap modulation; however, fully collapsing the gap is not always optimal, and intermediate, task-dependent separation yields the best results. These findings position the modality gap as a tunable property of multimodal representations rather than a quantity that should be universally minimized.

While individual components for AI agent memory exist in prior systems, their architectural synthesis and formal grounding remain underexplored. We present Kumiho, a graph-native cognitive memory architecture grounded in formal belief revision semantics. The structural primitives required for cognitive memory -- immutable revisions, mutable tag pointers, typed dependency edges, URI-based addressing -- are identical to those required for managing agent-produced work as versionable assets, enabling a unified graph-native architecture that serves both purposes. The central formal contribution is a correspondence between the AGM belief revision framework and the operational semantics of a property graph memory system, proving satisfaction of the basic AGM postulates (K*2--K*6) and Hansson's belief base postulates (Relevance, Core-Retainment). The architecture implements a dual-store model (Redis working memory, Neo4j long-term graph) with hybrid fulltext and vector retrieval. On LoCoMo (token-level F1), Kumiho achieves 0.565 overall F1 (n=1,986) including 97.5% adversarial refusal accuracy. On LoCoMo-Plus, a Level-2 cognitive memory benchmark testing implicit constraint recall, Kumiho achieves 93.3% judge accuracy (n=401); independent reproduction by the benchmark authors yielded results in the mid-80% range, still substantially outperforming all published baselines (best: Gemini 2.5 Pro, 45.7%). Three architectural innovations drive the results: prospective indexing (LLM-generated future-scenario implications indexed at write time), event extraction (structured causal events preserved in summaries), and client-side LLM reranking. The architecture is model-decoupled: switching the answer model from GPT-4o-mini (~88%) to GPT-4o (93.3%) improves end-to-end accuracy without pipeline changes, at a total evaluation cost of ~$14 for 401 entries.

Surgical co-management (SCM) is an evidence-based model in which hospitalists jointly manage medically complex perioperative patients alongside surgical teams. Despite its clinical and financial value, SCM is limited by the need to manually identify eligible patients. To determine whether SCM triage can be automated, we conducted a prospective, unblinded study at Stanford Health Care in which an LLM-based, electronic health record (EHR)-integrated triage tool (SCM Navigator) provided SCM recommendations followed by physician review. Using pre-operative documentation, structured data, and clinical criteria for perioperative morbidity, SCM Navigator categorized patients as appropriate, not appropriate, or possibly appropriate for SCM. Faculty indicated their clinical judgment and provided free-text feedback when they disagreed. Sensitivity, specificity, positive predictive value, and negative predictive value were measured using physician determinations as a reference. Free-text reasons were thematically categorized, and manual chart review was conducted on all false-negative cases and 30 randomly selected cases from the largest false-positive category. Since deployment, 6,193 cases have been triaged, of which 1,582 (23%) were recommended for hospitalist consultation. SCM Navigator displayed high sensitivity (0.94, 95% CI 0.91-0.96) and moderate specificity (0.74, 95% CI 0.71-0.77). Post-hoc chart review suggested most discrepancies reflect modifiable gaps in clinical criteria, institutional workflow, or physician practice variability rather than LLM misclassification, which accounted for 2 of 19 (11%) false-negative cases. These findings demonstrate that an LLM-powered, EHR-integrated, human-in-the-loop AI system can accurately and safely triage surgical patients for SCM, and that AI-enabled screening tools can augment and potentially automate time-intensive clinical workflows.

Auto-formalization (AF) translates natural-language reasoning problems into solver-executable programs, enabling symbolic solvers to perform sound logical deduction. In practice, however, AF pipelines are currently brittle: programs may fail to execute, or execute but encode incorrect semantics. While prior work largely mitigates syntactic failures via repairs based on solver feedback, reducing semantics failures remains a major bottleneck. We propose Draft-and-Prune (D&P), an inference-time framework that improves AF-based logical reasoning via diversity and verification. D&P first drafts multiple natural-language plans and conditions program generation on them. It further prunes executable but contradictory or ambiguous formalizations, and aggregates predictions from surviving paths via majority voting. Across four representative benchmarks (AR-LSAT, ProofWriter, PrOntoQA, LogicalDeduction), D&P substantially strengthens AF-based reasoning without extra supervision. On AR-LSAT, in the AF-only setting, D&P achieves 78.43% accuracy with GPT-4 and 78.00% accuracy with GPT-4o, significantly outperforming the strongest AF baselines MAD-LOGIC and CLOVER. D&P then attains near-ceiling performance on the other benchmarks, including 100% on PrOntoQA and LogicalDeduction.

Large audio-language models (LALMs) can produce expressive speech, yet reliable emotion control remains elusive: conversions often miss the target affect and may degrade linguistic fidelity through refusals, hallucinations, or paraphrase. We present, to our knowledge, the first neuron-level study of emotion control in speech-generative LALMs and demonstrate that compact emotion-sensitive neurons (ESNs) are causally actionable, enabling training-free emotion steering at inference time. ESNs are identified via success-filtered activation aggregation enforcing both emotion realization and content preservation. Across three LALMs (Qwen2.5-Omni-7B, MiniCPM-o 4.5, Kimi-Audio), ESN interventions yield emotion-specific gains that generalize to unseen speakers and are supported by automatic and human evaluation. Controllability depends on selector design, mask sparsity, filtering, and intervention strength. Our results establish a mechanistic framework for training-free emotion control in speech generation.

Kolmogorov-Arnold Networks (KANs) have gained attention for their potential to outperform Multi-Layer Perceptrons (MLPs) in terms of parameter efficiency and interpretability. Unlike traditional MLPs, KANs use learnable non-linear activation functions, typically spline functions, expressed as linear combinations of basis splines (B-splines). B-spline coefficients serve as the model's learnable parameters. However, evaluating these spline functions increases computational complexity during inference. Conventional quantization reduces this complexity by lowering the numerical precision of parameters and activations. However, the impact of quantization on KANs, and especially its effectiveness in reducing computational complexity, is largely unexplored, particularly for quantization levels below 8 bits. The study investigates the impact of low-bit quantization on KANs and its impact on computational complexity and hardware efficiency. Results show that B-splines can be quantized to 2-3 bits with negligible loss in accuracy, significantly reducing computational complexity. Hence, we investigate the potential of using low-bit quantized precomputed tables as a replacement for the recursive B-spline algorithm. This approach aims to further reduce the computational complexity of KANs and enhance hardware efficiency while maintaining accuracy. For example, ResKAN18 achieves a 50x reduction in BitOps without loss of accuracy using low-bit-quantized B-spline tables. Additionally, precomputed 8-bit lookup tables improve GPU inference speedup by up to 2.9x, while on FPGA-based systolic-array accelerators, reducing B-spline table precision from 8 to 3 bits cuts resource usage by 36%, increases clock frequency by 50%, and enhances speedup by 1.24x. On a 28nm FD-SOI ASIC, reducing the B-spline bit-width from 16 to 3 bits achieves 72% area reduction and 50% higher maximum frequency.

Multimodal Large Language Models (MLLMs) are increasingly applied to pixel-level vision tasks, yet their intrinsic capacity for spatial understanding remains poorly understood. We investigate segmentation capacity through a layerwise linear probing evaluation across the entire MLLM pipeline: vision encoder, adapter, and LLM. We further conduct an intervention based attention knockout analysis to test whether cross-token attention progressively refines visual representations, and an evaluation of bidirectional attention among image tokens on spatial consistency. Our analysis reveals that the adapter introduces a segmentation representation drop-off, but LLM layers progressively recover through attention-mediated refinement, where correctly classified tokens steer misclassified neighbors toward the correct label. At early image token positions, this recovery is bounded by causal attention, which bidirectional attention among image tokens alleviates. These findings provide a mechanistic account of how MLLMs process visual information for segmentation, informing the design of future segmentation-capable models.

The rapid proliferation of Large Language Models (LLMs) has created a profound digital divide, effectively excluding indigenous languages of the Global South from the AI revolution. The Tharu language, an Indo-Aryan vernacular spoken by approximately 1.7 million people across the Terai belt of Nepal and India, exemplifies this crisis. Despite a rich oral tradition, Tharu suffers from severe data scarcity and linguistic fragmentation, causing state-of-the-art multilingual models to routinely "hallucinate" or default to dominant high-resource neighbors like Hindi and Nepali due to contamination in pre-training corpora. This paper presents Tharu-LLaMA (3B), a specialized instruction-following model designed to address this exclusion. We introduce TharuChat, a novel dataset constructed via a LLM-to-Human bootstrapping pipeline. We utilized prompt-engineered Gemini models, fed with Rana Tharu grammar and folklore, to synthesize training data. Unlike curated gold-standard corpora, TharuChat reflects the noisy, heterogeneous linguistic reality of the region: it is predominantly anchored in Rana Tharu (~70%) while integrating elements of Dangaura and Kochila dialects. We provide a transparent analysis of the dataset's limitations, including dialectal code-mixing and residual Awadhi/Hindi influence. Through a rigorous empirical ablation study, we demonstrate that despite these imperfections, small-scale synthetic data is highly effective, increasing the dataset volume from 25% to 100% results in a linear reduction in perplexity from 6.42 to 2.88. The resulting model serves as a proof-of-concept for the preservation of under-resourced Himalayan languages via generative AI, achievable on consumer-grade hardware.

Multi-site neuroimaging analysis is fundamentally confounded by scanner-induced covariate shifts, where the marginal distribution of voxel intensities $P(\mathbf{x})$ varies non-linearly across acquisition protocols while the conditional anatomy $P(\mathbf{y}|\mathbf{x})$ remains constant. This is particularly detrimental to radiomic reproducibility, where acquisition variance often exceeds biological pathology variance. Existing statistical harmonization methods (e.g., ComBat) operate in feature space, precluding spatial downstream tasks, while standard deep learning approaches are theoretically bounded by local effective receptive fields (ERF), failing to model the global intensity correlations characteristic of field-strength bias. We propose SA-CycleGAN-2.5D, a domain adaptation framework motivated by the $HΔH$-divergence bound of Ben-David et al., integrating three architectural innovations: (1) A 2.5D tri-planar manifold injection preserving through-plane gradients $\nabla_z$ at $O(HW)$ complexity; (2) A U-ResNet generator with dense voxel-to-voxel self-attention, surpassing the $O(\sqrt{L})$ receptive field limit of CNNs to model global scanner field biases; and (3) A spectrally-normalized discriminator constraining the Lipschitz constant ($K_D \le 1$) for stable adversarial optimization. Evaluated on 654 glioma patients across two institutional domains (BraTS and UPenn-GBM), our method reduces Maximum Mean Discrepancy (MMD) by 99.1% ($1.729 \to 0.015$) and degrades domain classifier accuracy to near-chance (59.7%). Ablation confirms that global attention is statistically essential (Cohen's $d = 1.32$, $p < 0.001$) for the harder heterogeneous-to-homogeneous translation direction. By bridging 2D efficiency and 3D consistency, our framework yields voxel-level harmonized images that preserve tumor pathophysiology, enabling reproducible multi-center radiomic analysis.

Post-training alignment optimizes language models to match human preference signals, but this objective is not equivalent to modeling observed human behavior. We compare 120 base-aligned model pairs on more than 10,000 real human decisions in multi-round strategic games - bargaining, persuasion, negotiation, and repeated matrix games. In these settings, base models outperform their aligned counterparts in predicting human choices by nearly 10:1, robustly across model families, prompt formulations, and game configurations. This pattern reverses, however, in settings where human behavior is more likely to follow normative predictions: aligned models dominate on one-shot textbook games across all 12 types tested and on non-strategic lottery choices - and even within the multi-round games themselves, at round one, before interaction history develops. This boundary-condition pattern suggests that alignment induces a normative bias: it improves prediction when human behavior is relatively well captured by normative solutions, but hurts prediction in multi-round strategic settings, where behavior is shaped by descriptive dynamics such as reciprocity, retaliation, and history-dependent adaptation. These results reveal a fundamental trade-off between optimizing models for human use and using them as proxies for human behavior.

Responsible use of AI demands that we protect sensitive information without undermining the usefulness of data, an imperative that has become acute in the age of large language models. We address this challenge with an on-premise, LLM-driven substitution pipeline that anonymizes text by replacing personally identifiable information (PII) with realistic, type-consistent surrogates. Executed entirely within organizational boundaries using local LLMs, the approach prevents data egress while preserving fluency and task-relevant semantics. We conduct a systematic, multi-metric, cross-technique evaluation on the Action-Based Conversation Dataset, benchmarking against industry standards (Microsoft Presidio and Google DLP) and a state-of-the-art approach (ZSTS, in redaction-only and redaction-plus-substitution variants). Our protocol jointly measures privacy, semantic utility, and trainability under privacy via a lifecycle-ready criterion obtained by fine-tuning a compact encoder (BERT+LoRA) on sanitized text. In addition, we assess agentic Q&A performance by inserting an on-premise anonymization layer before the answering LLM and evaluating the quality of its responses. This intermediate, type-preserving substitution stage ensures that no sensitive content is exposed to third-party APIs, enabling responsible deployment of Q\&A agents without compromising confidentiality. Our method attains state-of-the-art privacy, minimal topical drift, strong factual utility, and low trainability loss, outperforming rule-based approaches and named-entity recognition (NER) baselines and ZSTS variants on the combined privacy--utility--trainability frontier. These results show that local LLM substitution yields anonymized corpora that are both responsible to use and operationally valuable: safe for agentic pipelines and suitable for downstream fine-tuning with limited degradation.

With the advent of AI agents, automatic scientific discovery has become a tenable goal. Many recent works scaffold agentic systems that can perform machine learning research, but don't offer a principled way to train such agents -- and current LLMs often generate plausible-looking but ineffective ideas. To make progress on training agents that can learn from doing, we provide a novel synthetic environment generation pipeline targeting machine learning agents. Our pipeline automatically synthesizes machine learning challenges compatible with the SWE-agent framework, covering topic sampling, dataset proposal, and code generation. The resulting synthetic tasks are 1) grounded in real machine learning datasets, because the proposed datasets are verified against the Huggingface API and are 2) verified for higher quality with a self-debugging loop. To validate the effectiveness of our synthetic tasks, we tackle MLGym, a benchmark for machine learning tasks. From the synthetic tasks, we sample trajectories from a teacher model (GPT-5), then use the trajectories to train a student model (Qwen3-4B and Qwen3-8B). The student models trained with our synthetic tasks achieve improved performance on MLGym, raising the AUP metric by 9% for Qwen3-4B and 12% for Qwen3-8B.

When organizations delegate text generation tasks to AI providers via pay-for-performance contracts, expected payments rise when evaluation is noisy. As evaluation methods become more elaborate, the economic benefits of decreased noise are often overshadowed by increased evaluation costs. In this work, we introduce adaptive contracts for AI delegation, which allow detailed evaluation to be performed selectively after observing an initial coarse signal in order to conserve resources. We make three sets of contributions: First, we provide efficient algorithms for computing optimal adaptive contracts under natural assumptions or when core problem dimensions are small, and prove hardness of approximation in the general unstructured case. We then formulate alternative models of randomized adaptive contracts and discuss their benefits and limitations. Finally, we empirically demonstrate the benefits of adaptivity over non-adaptive baselines using question-answering and code-generation datasets.

Efficient and accurate prediction of Multiphysics evolution across diverse cell geometries is fundamental to the design, management and safety of lithium-ion batteries. However, existing computational frameworks struggle to capture the coupled electrochemical, thermal, and mechanical dynamics across diverse cell geometries and varying operating conditions. Here, we present a Neural Bundle Map (NBM), a mathematically rigorous framework that reformulates multiphysics evolution as a bundle map over a geometric base manifold. This approach enables the complete decoupling of geometric complexity from underlying physical laws, ensuring strong operator continuity across varying domains. Our framework achieves high-fidelity spatiotemporal predictions with a normalized mean absolute error of less than 1% across varying configurations, while maintaining stability during long-horizon forecasting far beyond the training window and reducing computational costs by two orders of magnitude compared with conventional solvers. Leveraging this capability, we rapidly explored a vast configurational space to identify an optimal battery design that yields a 38% increase in energy density while adhering to thermal safety constraints. Furthermore, the NBM demonstrates remarkable scalability to multi-cell systems through few-shot transfer learning, providing a foundational paradigm for the intelligent design and real-time monitoring of complex energy storage infrastructures.

Register-Transfer Level (RTL) synthesis and summarization are central to hardware design automation but remain challenging for Large Language Models (LLMs) due to rigid HDL syntax, limited supervision, and weak alignment with natural language. Existing prompting and retrieval-augmented generation (RAG) methods have not incorporated symbolic planning, limiting their structural precision. We introduce SYMDIREC, a neuro-symbolic framework that decomposes RTL tasks into symbolic subgoals, retrieves relevant code via a fine-tuned retriever, and assembles verified outputs through LLM reasoning. Supporting both Verilog and VHDL without LLM fine-tuning, SYMDIREC achieves ~20% higher Pass@1 rates for synthesis and 15-20% ROUGE-L improvements for summarization over prompting and RAG baselines, demonstrating the benefits of symbolic guidance in RTL tasks.

Domain-specific finetuning is essential for dense retrievers, yet not all training pairs contribute equally to the learning process. We introduce OPERA, a data pruning framework that exploits this heterogeneity to improve both the effectiveness and efficiency of retrieval model adaptation. We first investigate static pruning (SP), which retains only high-similarity query-document pairs, revealing an intrinsic quality-coverage tradeoff: ranking (NDCG) improves while retrieval (Recall) can degrade due to reduced query diversity. To resolve this tradeoff, we propose a two-stage dynamic pruning (DP) strategy that adaptively modulates sampling probabilities at both query and document levels throughout training, prioritizing high-quality examples while maintaining access to the full training set. Evaluations across eight datasets spanning six domains demonstrate the effectiveness of both approaches: SP improves ranking over standard finetuning (NDCG@10 +0.5\%), while DP achieves the strongest performance on both ranking (NDCG@10 +1.9\%) and retrieval (Recall@20 +0.7\%), with an average rank of 1.38 across all methods. These findings scale to Qwen3-Embedding, an LLM-based dense retriever, confirming architecture-agnostic benefits. Notably, DP reaches comparable performance in less than 50\% of the training time required by standard finetuning.

Optimizing Register Transfer Level (RTL) code is a critical step in Electronic Design Automation (EDA) for improving power, performance, and area (PPA). We present CODMAS (Collaborative Optimization via a Dialectic Multi-Agent System), a framework that combines structured dialectic reasoning with domain-aware code generation and deterministic evaluation to automate RTL optimization. At the core of CODMAS are two dialectic agents: the Articulator, inspired by rubber-duck debugging, which articulates stepwise transformation plans and exposes latent assumptions; and the Hypothesis Partner, which predicts outcomes and reconciles deviations between expected and actual behavior to guide targeted refinements. These agents direct a Domain-Specific Coding Agent (DCA) to generate architecture-aware Verilog edits and a Code Evaluation Agent (CEA) to verify syntax, functionality, and PPA metrics. We introduce RTLOPT, a benchmark of 120 Verilog triples (unoptimized, optimized, testbench) for pipelining and clock-gating transformations. Across proprietary and open LLMs, CODMAS achieves ~25% reduction in critical path delay for pipelining and ~22% power reduction for clock gating, while reducing functional and compilation failures compared to strong prompting and agentic baselines. These results demonstrate that structured multi-agent reasoning can significantly enhance automated RTL optimization and scale to more complex designs and broader optimization tasks.

Large language models (LLMs) can produce chains of thought (CoT) that do not accurately reflect the actual factors driving their answers. In multiple-choice settings with an injected hint favoring a particular option, models may shift their final answer toward the hinted option and produce a CoT that rationalizes the response without acknowledging the hint - an instance of motivated reasoning. We study this phenomenon across multiple LLM families and datasets demonstrating that motivated reasoning can be identified by probing internal activations even in cases when it cannot be easily determined from CoT. Using supervised probes trained on the model's residual stream, we show that (i) pre-generation probes, applied before any CoT tokens are generated, predict motivated reasoning as well as a LLM-based CoT monitor that accesses the full CoT trace, and (ii) post-generation probes, applied after CoT generation, outperform the same monitor. Together, these results show that motivated reasoning is detected more reliably from internal representations than from CoT monitoring. Moreover, pre-generation probing can flag motivated behavior early, potentially avoiding unnecessary generation.

The real world is non-stationary and infinitely complex, requiring intelligent agents to learn continually without the prohibitive cost of retraining from scratch. While online continual learning offers a framework for this setting, learning new information often interferes with previously acquired knowledge, causes forgetting and degraded generalization. To address this, we propose Abstraction-Augmented Training (AAT), a loss-level modification encouraging models to capture the latent relational structure shared across examples. By jointly optimizing over concrete instances and their abstract representations, AAT introduces a memory-efficient inductive bias that stabilizes learning in strictly online data streams, eliminating the need for a replay buffer. To capture the multi-faceted nature of abstraction, we introduce and evaluate AAT on two benchmarks: a controlled relational dataset where abstraction is realized through entity masking, and a narrative dataset where abstraction is expressed through shared proverbs. Our results show that AAT achieves performance comparable to or exceeding strong experience replay (ER) baselines, despite requiring zero additional memory and only minimal changes to the training objective. This work highlights structural abstraction as a powerful, memory-free alternative to ER.

The success of large-scale pretraining in NLP and computer vision has catalyzed growing efforts to develop analogous foundation models for the physical sciences. However, pretraining strategies using atomistic data remain underexplored. To date, large-scale supervised pretraining on DFT force-energy labels has provided the strongest performance gains to downstream property prediction, out-performing existing methods of self-supervised learning (SSL) which remain limited to ground-state geometries, and/or single domains of atomistic data. We address these shortcomings with Self-Conditioned Denoising (SCD), a backbone-agnostic reconstruction objective that utilizes self-embeddings for conditional denoising across any domain of atomistic data, including small molecules, proteins, periodic materials, and 'non-equilibrium' geometries. When controlled for backbone architecture and pretraining dataset, SCD significantly outperforms previous SSL methods on downstream benchmarks and matches or exceeds the performance of supervised force-energy pretraining. We show that a small, fast GNN pretrained by SCD can achieve competitive or superior performance to larger models pretrained on significantly larger labeled or unlabeled datasets, across tasks in multiple domains. Our code is available at: https://github.com/TyJPerez/SelfConditionedDenoisingAtoms

Formal specifications, such as pre- and post-conditions provide a solid basis for performing thorough program verification. However, developers rarely provide such formal specifications, hence if AI could help in constructing them, it would make formal verification possible or at least make automated testing much more effective. This paper presents a study on the ability of LLMs in generating formal FULL pre- and post-conditions of a program, given its natural language description. 24 state-of-the-art LLMs were evaluated on a freshly prepared dataset of 40 tasks. The paper investigates specifications of varying difficulty and discusses a set of more refined performance metrics in addition the general accept@1 performance. It also investigates the impact of using automatically generated tests for validation of the solutions proposed by LLMs. The results of the experiment reveal that, in general LLMs can produce valid pre- and post-conditions based on natural language descriptions of programs. Incorrect solutions from proprietary models are also often near correct. A closer inspection shows that open-source models tend to result in a higher proportion of erroneous results while proprietary models tend to have slightly higher false negative rates. Interestingly, the results also show that augmenting the manually prepared dataset with automatically generated tests leads to the exposure of wrong solutions, which would have otherwise been accepted as correct. In general, LLMs perform better in formalizing pre-conditions than on post-conditions and proprietary models perform better than open ones. However, none of the LLMs were able to correctly formalize all the tasks in our benchmark. Overall, the effectiveness and reliability of LLMs in formalizing pre- and post-conditions could be greatly improved by using a good test suite that checks the correctness of the LLM generated formalizations.

Accurate diagnosis of Alzheimer's disease (AD) requires handling tabular biomarker data, yet such data are often small and incomplete, where deep learning models frequently fail to outperform classical methods. Pretrained large language models (LLMs) offer few-shot generalization, structured reasoning, and interpretable outputs, providing a powerful paradigm shift for clinical prediction. We propose TAP-GPT Tabular Alzheimer's Prediction GPT, a domain-adapted tabular LLM framework built on TableGPT2 and fine-tuned for few-shot AD classification using tabular prompts rather than plain texts. We evaluate TAP-GPT across four ADNI-derived datasets, including QT-PAD biomarkers and region-level structural MRI, amyloid PET, and tau PET for binary AD classification. Across multimodal and unimodal settings, TAP-GPT improves upon its backbone models and outperforms traditional machine learning baselines in the few-shot setting while remaining competitive with state-of-the-art general-purpose LLMs. We show that feature selection mitigates degradation in high-dimensional inputs and that TAP-GPT maintains stable performance under simulated and real-world missingness without imputation. Additionally, TAP-GPT produces structured, modality-aware reasoning aligned with established AD biology and shows greater stability under self-reflection, supporting its use in iterative multi-agent systems. To our knowledge, this is the first systematic application of a tabular-specialized LLM to multimodal biomarker-based AD prediction, demonstrating that such pretrained models can effectively address structured clinical prediction tasks and laying the foundation for tabular LLM-driven multi-agent clinical decision-support systems. The source code is publicly available on GitHub: https://github.com/sophie-kearney/TAP-GPT.

Large language model (LLM) agents are increasingly used for complex tasks, yet deployed agents often remain static, failing to adapt as user needs evolve. This creates a tension between the need for continuous service and the necessity of updating capabilities to match shifting task distributions. On platforms like OpenClaw, which handle diverse workloads across 20+ channels, existing methods either store raw trajectories without distilling knowledge, maintain static skill libraries, or require disruptive downtime for retraining. We present MetaClaw, a continual meta-learning framework that jointly evolves a base LLM policy and a library of reusable behavioral skills. MetaClaw employs two complementary mechanisms. Skill-driven fast adaptation analyzes failure trajectories via an LLM evolver to synthesize new skills, enabling immediate improvement with zero downtime. Opportunistic policy optimization performs gradient-based updates via cloud LoRA fine-tuning and Reinforcement Learning with a Process Reward Model (RL-PRM). This is triggered during user-inactive windows by the Opportunistic Meta-Learning Scheduler (OMLS), which monitors system inactivity and calendar data. These mechanisms are mutually reinforcing: a refined policy generates better trajectories for skill synthesis, while richer skills provide higher-quality data for policy optimization. To prevent data contamination, a versioning mechanism separates support and query data. Built on a proxy-based architecture, MetaClaw scales to production-size LLMs without local GPUs. Experiments on MetaClaw-Bench and AutoResearchClaw show that skill-driven adaptation improves accuracy by up to 32% relative. The full pipeline advances Kimi-K2.5 accuracy from 21.4% to 40.6% and increases composite robustness by 18.3%. Code is available at https://github.com/aiming-lab/MetaClaw.

Retrieval augmented generation systems have become an integral part of everyday life. Whether in internet search engines, email systems, or service chatbots, these systems are based on context retrieval and answer generation with large language models. With their spread, also the security vulnerabilities increase. Attackers become increasingly focused on these systems and various hacking approaches are developed. Manipulating the context documents is a way to persist attacks and make them affect all users. Therefore, detecting compromised, adversarial context documents early is crucial for security. While supervised approaches require a large amount of labeled adversarial contexts, we propose an unsupervised approach, being able to detect also zero day attacks. We conduct a preliminary study to show appropriate indicators for adversarial contexts. For that purpose generator activations, output embeddings, and an entropy-based uncertainty measure turn out as suitable, complementary quantities. With an elementary statistical outlier detection, we propose and compare their detection abilities. Furthermore, we show that the target prompt, which the attacker wants to manipulate, is not required for a successful detection. Moreover, our results indicate that a simple context summary generation might even be superior in finding manipulated contexts.

Explainable Boosting Machines (EBMs) provide transparent predictions through additive shape functions, enabling direct inspection of feature contributions. However, EBMs can learn non-physical relationships that reduce their reliability in natural hazard applications. This study presents a domain-informed framework to improve the physical consistency of EBMs for lateral spreading prediction. Our approach modifies learned shape functions based on domain knowledge. These modifications correct non-physical behavior while maintaining data-driven patterns. We apply the method to the 2011 Christchurch earthquake dataset and correct non-physical trends observed in the original EBM. The resulting model produces more physically consistent global and local explanations, with an acceptable tradeoff in accuracy (4--5\%).

Code generation large language models (LLMs) are increasingly integrated into modern software development workflows. Recent work has shown that these models are vulnerable to backdoor and poisoning attacks that induce the generation of insecure code, yet effective defenses remain limited. Existing scanning approaches rely on token-level generation consistency to invert attack targets, which is ineffective for source code where identical semantics can appear in diverse syntactic forms. We present CodeScan, which, to the best of our knowledge, is the first poisoning-scanning framework tailored to code generation models. CodeScan identifies attack targets by analyzing structural similarities across multiple generations conditioned on different clean prompts. It combines iterative divergence analysis with abstract syntax tree (AST)-based normalization to abstract away surface-level variation and unify semantically equivalent code, isolating structures that recur consistently across generations. CodeScan then applies LLM-based vulnerability analysis to determine whether the extracted structures contain security vulnerabilities and flags the model as compromised when such a structure is found. We evaluate CodeScan against four representative attacks under both backdoor and poisoning settings across three real-world vulnerability classes. Experiments on 108 models spanning three architectures and multiple model sizes demonstrate 97%+ detection accuracy with substantially lower false positives than prior methods.

Iris presentation attack detection (PAD) is critical for secure biometric deployments, yet developing specialized models faces significant practical barriers: collecting data representing future unknown attacks is impossible, and collecting diverse-enough data, yet still limited in terms of its predictive power, is expensive. Additionally, sharing biometric data raises privacy concerns. Due to rapid emergence of new attack vectors demanding adaptable solutions, we thus investigate in this paper whether general-purpose multimodal large language models (MLLMs) can perform iris PAD when augmented with human expert knowledge, operating under strict privacy constraints that prohibit sending biometric data to public cloud MLLM services. Through analysis of vision encoder embeddings applied to our dataset, we demonstrate that pre-trained vision transformers in MLLMs inherently cluster many iris attack types despite never being explicitly trained for this task. However, where clustering shows overlap between attack classes, we find that structured prompts incorporating human salience (verbal descriptions from subjects identifying attack indicators) enable these models to resolve ambiguities. Testing on an IRB-restricted dataset of 224 iris images spanning seven attack types, using only university-approved services (Gemini 2.5 Pro) or locally-hosted models (e.g., Llama 3.2-Vision), we show that Gemini with expert-informed prompts outperforms both a specialized convolutional neural networks (CNN)-based baseline and human examiners, while the locally-deployable Llama achieves near-human performance. Our results establish that MLLMs deployable within institutional privacy constraints offer a viable path for iris PAD.

Large language models (LLMs) are increasingly used as automated judges and synthetic labelers, especially in low-label settings. Yet these systems are stochastic and often overconfident, which makes deployment decisions difficult when external ground truth is limited. We propose a practical calibration protocol based on controlled input interventions: if noise severity increases, task performance should exhibit a statistically significant deterioration trend. We operationalize this with a slope-based hypothesis test over repeated trials, using signal-to-noise-ratio (SNR) perturbations for tabular data and lexical perturbations for text data. Across UCI tabular benchmarks and four text classification datasets, we find clear modality-dependent behavior. Our results reveal a modality gap: while text-based judges degrade predictably, the majority of tabular datasets show a lack of statistically significant performance deterioration even under significant signal-to-noise reduction. Interestingly we find that model performance is lower on datasets that are insensitive to noise interventions. We present a reproducible methodology and reporting protocol for robust LLM-judge calibration under distribution shift.

Evaluating the grammatical competence of second language (L2) learners is essential both for providing targeted feedback and for assessing proficiency. To achieve this, we propose a novel framework leveraging the English Grammar Profile (EGP), a taxonomy of grammatical constructs mapped to the proficiency levels of the Common European Framework of Reference (CEFR), to detect learners' attempts at grammatical constructs and classify them as successful or unsuccessful. This detection can then be used to provide fine-grained feedback. Moreover, the grammatical constructs are used as predictors of proficiency assessment by using automatically detected attempts as predictors of holistic CEFR proficiency. For the selection of grammatical constructs derived from the EGP, rule-based and LLM-based classifiers are compared. We show that LLMs outperform rule-based methods on semantically and pragmatically nuanced constructs, while rule-based approaches remain competitive for constructs that rely purely on morphological or syntactic features and do not require semantic interpretation. For proficiency assessment, we evaluate both rule-based and hybrid pipelines and show that a hybrid approach combining a rule-based pre-filter with an LLM consistently yields the strongest performance. Since our framework operates on pairs of original learner sentences and their corrected counterparts, we also evaluate a fully automated pipeline using automatic grammatical error correction. This pipeline closely approaches the performance of semi-automated systems based on manual corrections, particularly for the detection of successful attempts at grammatical constructs. Overall, our framework emphasises learners' successful attempts in addition to unsuccessful ones, enabling positive, formative feedback and providing actionable insights into grammatical development.

The emerging agentic web envisions AI agents that reliably fulfill users' natural-language (NL)-based tasks by interacting with existing web services. However, existing authorization models are misaligned with this vision. In particular, today's operator-scoped authorization, exemplified by OAuth, grants broad permissions tied to operators (e.g., the transfer operator) rather than to the specific operations (e.g., transfer $100 to Bob) implied by a user's task. This will inevitably result in overprivileged agents. We introduce Precise Task-Scoped Implicit Authorization (PAuth), a fundamentally different model in which submitting an NL task implicitly authorizes only the concrete operations required for its faithful execution. To make this enforceable at servers, we propose NL slices: symbolic specifications of the calls each service expects, derived from the task and upstream results. Complementing this, we also propose envelopes: special data structure to bind each operand's concrete value to its symbolic provenance, enabling servers to verify that all operands arise from legitimate computations. PAuth is prototyped in the agent-security evaluation framework AgentDojo. We evaluate it in both benign settings and attack scenarios where a spurious operation is injected into an otherwise normal task. In all benign tests, PAuth executes the tasks successfully without requiring any additional permissions. In all attack tests, PAuth correctly raises warnings about missing permissions. These results demonstrate that PAuth's reasoning about permissions is indeed precise. We further analyze the characteristics of these tasks and measure the associated token costs.

Deducing whodunit proves challenging for LLM agents. In this paper, we implement a text-based multi-agent version of the classic board game Clue as a rule-based testbed for evaluating multi-step deductive reasoning, with six agents drawn from GPT-4o-mini and Gemini-2.5-Flash. We further investigate whether fine-tuning on structured logic puzzles transfers to improved in-game reasoning and gameplay. Across 18 simulated games, agents achieve only four correct wins, indicating difficulty in maintaining consistent deductive reasoning over the course of a full game. Additionally, we find that fine-tuning does not reliably improve performance and, in some cases, appears to increase reasoning volume without improving reasoning precision.

The growing energy demands of computational systems necessitate a fundamental shift from performance-centric design to one that treats energy consumption as one of the primary design considerations. Current approaches treat energy consumption as an aggregate, deterministic property, overlooking the path-dependent nature of computation, where different execution paths through the same software consume dramatically different energy. We introduce the Energy Flow Graph (EFG), a formal model that represents computational processes as state-transition systems with energy costs for both states and transitions. EFG enables various applications in software engineering, including static analysis of energy-optimal execution paths and a multiplicative cascade model that predicts combined optimization effects without exhaustive testing. Our early experiments demonstrate EFG's versatility across domains: in software programs validated through 3.5 million executions, 15.6% of solutions exhibit high path-dependent variance (CV $>$ 0.1), while structural optimization reveals up to 705$\times$ energy reduction. In AI pipelines, the cascade model predicts optimization combinations within 5.1% error, enabling selection from 4.2 million possibilities using only 22 measurements. The EFG transforms energy optimization from trial-and-error to systematic analysis, providing a foundation for green software engineering across computational domains.

We introduce a general framework for analyzing learning algorithms based on the notion of self-regularization, which captures implicit complexity control without requiring explicit regularization. This is motivated by previous observations that many algorithms, such as gradient-descent based learning, exhibit implicit regularization. In a nutshell, for a self-regularized algorithm the complexity of the predictor is inherently controlled by that of the simplest comparator achieving the same empirical risk. This framework is sufficiently rich to cover both classical regularized empirical risk minimization and gradient descent. Building on self-regularization, we provide a thorough statistical analysis of such algorithms including minmax-optimal rates, where it suffices to show that the algorithm is self-regularized -- all further requirements stem from the learning problem itself. Finally, we discuss the problem of data-dependent hyperparameter selection, providing a general result which yields minmax-optimal rates up to a double logarithmic factor and covers data-driven early stopping for RKHS-based gradient descent.

Reinforcement Learning (RL) has shown promise in various robotics applications, yet its deployment on real systems is still limited due to safety and operational constraints. The safe RL field has gained considerable attention in recent years, which focuses on imposing safety constraints throughout the learning process. However, real systems often require more complex constraints than just safety, such as periodic recharging or time-bounded visits to specific regions. Imposing such spatio-temporal tasks during learning still remains a challenge. Signal Temporal Logic (STL) is a formal language for specifying temporal properties of real-valued signals and provides a way to express such complex tasks. In this paper, we propose a framework that leverages sequential control barrier functions and model-free RL to ensure that the given STL tasks are satisfied throughout the learning process. Our method extends beyond traditional safety constraints by enforcing rich STL specifications, which can involve visits to dynamic targets with unknown trajectories. We also demonstrate the effectiveness of our framework through various simulations.

Agentic AI systems can now generate code with remarkable fluency, but a fundamental question remains: \emph{does the generated code actually do what the user intended?} The gap between informal natural language requirements and precise program behavior -- the \emph{intent gap} -- has always plagued software engineering, but AI-generated code amplifies it to an unprecedented scale. This article argues that \textbf{intent formalization} -- the translation of informal user intent into a set of checkable formal specifications -- is the key challenge that will determine whether AI makes software more reliable or merely more abundant. Intent formalization offers a tradeoff spectrum suitable to the reliability needs of different contexts: from lightweight tests that disambiguate likely misinterpretations, through full functional specifications for formal verification, to domain-specific languages from which correct code is synthesized automatically. The central bottleneck is \emph{validating specifications}: since there is no oracle for specification correctness other than the user, we need semi-automated metrics that can assess specification quality with or without code, through lightweight user interaction and proxy artifacts such as tests. We survey early research that demonstrates the \emph{potential} of this approach: interactive test-driven formalization that improves program correctness, AI-generated postconditions that catch real-world bugs missed by prior methods, and end-to-end verified pipelines that produce provably correct code from informal specifications. We outline the open research challenges -- scaling beyond benchmarks, achieving compositionality over changes, metrics for validating specifications, handling rich logics, designing human-AI specification interactions -- that define a research agenda spanning AI, programming languages, formal methods, and human-computer interaction.

Personalized fall detection models can significantly improve accuracy by adapting to individual motion patterns, yet their effectiveness is often limited by the scarcity of real-world fall data and the dominance of non-fall feedback samples. This imbalance biases the model toward routine activities and weakens its sensitivity to true fall events. To address this challenge, we propose a personalization framework that combines semi-supervised clustering with contrastive learning to identify and balance the most informative user feedback samples. The framework is evaluated under three retraining strategies, including Training from Scratch (TFS), Transfer Learning (TL), and Few-Shot Learning (FSL), to assess adaptability across learning paradigms. Real-time experiments with ten participants show that the TFS approach achieves the highest performance, with up to a 25% improvement over the baseline, while FSL achieves the second-highest performance with a 7% improvement, demonstrating the effectiveness of selective personalization for real-world deployment.

Wikipedia is a critical source of information for millions of users across the Web. It serves as a key resource for large language models, search engines, question-answering systems, and other Web-based applications. In Wikipedia, content needs to be verifiable, meaning that readers can check that claims are backed by references to reliable sources. This depends on manual verification by editors, an effective but labor-intensive process, especially given the high volume of daily edits. To address this challenge, we introduce a multilingual machine learning system to assist editors in identifying claims requiring citations. Our approach is tested in 10 language editions of Wikipedia, outperforming existing benchmarks for reference need assessment. We not only consider machine learning evaluation metrics but also system requirements, allowing us to explore the trade-offs between model accuracy and computational efficiency under real-world infrastructure constraints. We deploy our system in production and release data and code to support further research.

Large language models (LLMs) are increasingly deployed as automated evaluators that assign numeric scores to model outputs, a paradigm known as LLM-as-a-Judge. However, standard Reinforcement Learning (RL) methods typically rely on binary rewards (e.g., 0-1 accuracy), thereby ignoring the ordinal structure inherent in regression tasks; for instance, they fail to recognize that predicting 4 is significantly better than predicting 1 when the ground truth is 5. Conversely, existing regression-aware approaches are often confined to Supervised Fine-Tuning (SFT), limiting their ability to explore optimal reasoning paths. To bridge this gap, we propose \textbf{REAL} (\underline{RE}gression-\underline{A}ware Reinforcement \underline{L}earning), a principled RL framework designed to optimize regression rewards, and also proven to be optimal for correlation metrics. A key technical challenge is that the regression objective is explicitly policy-dependent, thus invalidating standard policy gradient methods. To address this, we employ the generalized policy gradient estimator, which naturally decomposes optimization into two complementary components: (1) exploration over Chain-of-Thought (CoT) trajectory, and (2) regression-aware prediction refinement of the final score. Extensive experiments across model scales (8B to 32B) demonstrate that REAL consistently outperforms both regression-aware SFT baselines and standard RL methods, exhibiting significantly better generalization on out-of-domain benchmarks. On Qwen3-32B specifically, we achieve gains of +8.40 Pearson and +7.20 Spearman correlation over the SFT baseline, and +18.30/+11.20 over the base model. These findings highlight the critical value of integrating regression objectives into RL exploration for accurate LLM evaluation.

Decision-making problems often feature uncertainty stemming from heterogeneous and context-dependent human preferences. To address this, we propose a sequential learning-and-optimization pipeline to learn preference distributions and leverage them to solve downstream problems, for example risk-averse formulations. We focus on human choice settings that can be formulated as (integer) linear programs. In such settings, existing inverse optimization and choice modelling methods infer preferences from observed choices but typically produce point estimates or fail to capture contextual shifts, making them unsuitable for risk-averse decision-making. Using a bounded-variance score function gradient estimator, we train a predictive model mapping contextual features to a rich class of parameterizable distributions. This approach yields a maximum likelihood estimate. The model generates scenarios for unseen contexts in the subsequent optimization phase. In a synthetic ridesharing environment, our approach reduces average post-decision surprise by up to 114$\times$ compared to a risk-neutral approach with perfect predictions and up to 25$\times$ compared to leading risk-averse baselines.

Identity-based software signing tools aim to make software artifact provenance verifiable while reducing the operational burden of long-lived key management. However, there is limited cross-tool longitudinal evidence about which usability problems arise in practice and how those problems evolve as tools mature. This gap matters because unusable signing and verification workflows can lead to incomplete adoption, misconfiguration, or skipped verification, undermining intended integrity guarantees. We conducted the first mining-software-repositories study of five open-source identity-based signing ecosystems: Sigstore, OpenPubKey, HashiCorp Vault, Keyfactor, and Notary v2. We analyzed approximately 3,900 GitHub issues from Nov. 2021 to Nov. 2025. We coded each issue for the reported usability concern and the implicated architectural component, and compared patterns across tools and over time. Across ecosystems, reported concerns concentrate in verification workflows, policy and configuration surfaces, and integration boundaries. Longitudinal Poisson trend analysis shows substantial declines in reported issues for most ecosystems. However, across usability themes, workflow- and documentation-related concerns decline unevenly across tools and concern types, and verification workflows and configuration surfaces remain persistent friction points. These results indicate that identity-based signing reduces some usability burdens while relocating complexity to verification semantics, policy configuration, and deployment integration. Designing future signing ecosystems therefore requires treating verification semantics and release workflows as first-class usability targets rather than peripheral integration concerns.

Many wireless vision applications, such as autonomous driving, require preservation of global structural information rather than only per-pixel fidelity. However, existing Deep joint source-channel coding (DeepJSCC) schemes mainly optimize pixel-wise losses and provide no explicit protection of connectivity or topology. This letter proposes TopoJSCC, a topology-aware DeepJSCC framework that integrates persistent-homology regularizers to end-to-end training. Specifically, we enforce topological consistency by penalizing Wasserstein distances between cubical persistence diagrams of original and reconstructed images, and between Vietoris--Rips persistence of latent features before and after the channel to promote a robust latent manifold. TopoJSCC is based on end-to-end learning and requires no side information. Experiments show improved topology preservation and peak signal-to-noise ratio (PSNR) in low signal-to-noise ratio (SNR) and bandwidth-ratio regimes.

Large Language Models increasingly power critical infrastructure from healthcare to finance, yet their vulnerability to adversarial manipulation threatens system integrity and user safety. Despite growing deployment, no comprehensive comparative security assessment exists across major LLM architectures, leaving organizations unable to quantify risk or select appropriately secure LLMs for sensitive applications. This research addresses this gap by establishing a standardized vulnerability assessment framework and developing a multi-layered defensive system to protect against identified threats. We systematically evaluate five widely-deployed LLM families GPT-4, GPT-3.5 Turbo, Claude-3 Haiku, LLaMA-2-70B, and Gemini-2.5-pro against 10,000 adversarial prompts spanning six attack categories. Our assessment reveals critical security disparities, with vulnerability rates ranging from 11.9\% to 29.8\%, demonstrating that LLM capability does not correlate with security robustness. To mitigate these risks, we develop a production-ready defensive framework achieving 83\% average detection accuracy with only 5\% false positives. These results demonstrate that systematic security assessment combined with external defensive measures provides a viable path toward safer LLM deployment in production environments.

A common architectural pattern in advanced AI reasoning systems is the symbolic graph network: specialized agents or modules connected by delegation edges, routing tasks through a dynamic execution graph. Current schedulers optimize load and fitness but are geometry-blind: they do not model how failures propagate differently in tree-like versus cyclic regimes. In tree-like delegation, a single failure can cascade exponentially; in dense cyclic graphs, failures tend to self-limit. We identify this observability gap, quantify its system-level cost, and propose a lightweight mitigation. We formulate online geometry control for route-risk estimation on time-indexed execution graphs with route-local failure history. Our approach combines (i) a Euclidean spatio-temporal propagation baseline, (ii) a hyperbolic route-risk model with temporal decay (and optional burst excitation), and (iii) a learned geometry selector over structural features. The selector is a compact MLP (9->12->1) using six topology statistics plus three geometry-aware signals: BFS shell-growth slope, cycle-rank norm, and fitted Poincare curvature. On the Genesis 3 benchmark distribution, adaptive switching improves win rate in the hardest non_tree regime from 64-72% (fixed hyperbolic variants) to 92%, and achieves 87.2% overall win rate. To measure total system value, we compare against Genesis 3 routing without any spatio-temporal sidecar, using only native bandit/LinUCB signals (team fitness and mean node load). This baseline achieves 50.4% win rate overall and 20% in tree-like regimes; the full sidecar recovers 87.2% overall (+36.8 pp), with +48 to +68 pp gains in tree-like settings, consistent with a cascade-sensitivity analysis. Overall, a 133-parameter sidecar substantially mitigates geometry-blind failure propagation in one high-capability execution-graph system.

Ensembling Vision-Language Models (VLMs) from different providers maximizes benchmark accuracy, yet models from the same architectural family share correlated errors that standard voting ignores. We study this structure across 17 VLMs from 8 families on VQAv2, TextVQA, and GQA. Family-correlated errors reduce effective ensemble dimensionality to 2.5-3.6 independent voters and create a Misleading tier (1.5-6.5% of questions) where correlated majority errors destroy accuracy to 0% despite the best model being correct. We propose three family-aware methods. Hierarchical Family Voting (HFV) aggregates within families before voting across them, recovering +18-26 pp on the Misleading tier. QualRCCV, a training-free method weighting models by calibration, family quality, and inverse family size, is the first to beat calibrated voting on all three benchmarks (p<0.05). Learned Candidate Scoring (LCS) trains a cross-validated classifier to re-rank candidate answers using support breadth, family diversity, and model quality, achieving the largest gains: +0.68% VQAv2, +0.61% TextVQA, +2.45% GQA -- all significant -- and is the only learned method that never degrades any benchmark. On VQAv2 test-standard (EvalAI), LCS reaches 87.83% with 12 models, confirming generalization.

Image segmentation relies on large annotated datasets, which are expensive and slow to produce. Silver-standard (AI-generated) labels are easier to obtain, but they risk introducing bias. Self-supervised learning, needing only images, has become key for pre-training. Recent work combining contrastive learning with counterfactual generation improves representation learning for classification but does not readily extend to pixel-level tasks. We propose a pipeline combining counterfactual generation with dense contrastive learning via Dual-View (DVD-CL) and Multi-View (MVD-CL) methods, along with supervised variants that utilize available silver-standard annotations. A new visualisation algorithm, the Color-coded High Resolution Overlay map (CHRO-map) is also introduced. Experiments show annotation-free DVD-CL outperforms other dense contrastive learning methods, while supervised variants using silver-standard labels outperform training on the silver-standard labeled data directly, achieving $\sim$94% DSC on challenging data. These results highlight that pixel-level contrastive learning, enhanced by counterfactuals and silver-standard annotations, improves robustness to acquisition and pathological variations.

Decoding brain activity into natural language is a major challenge in AI with important applications in assistive communication, neurotechnology, and human-computer interaction. Most existing Brain-Computer Interface (BCI) approaches rely on memory-intensive fine-tuning of Large Language Models (LLMs) or encoder-decoder models on raw EEG signals, resulting in expensive training pipelines, limited accessibility, and potential exposure of sensitive neural data. We introduce SENSE (SEmantic Neural Sparse Extraction), a lightweight and privacy-preserving framework that translates non-invasive electroencephalography (EEG) into text without LLM fine-tuning. SENSE decouples decoding into two stages: on-device semantic retrieval and prompt-based language generation. EEG signals are locally mapped to a discrete textual space to extract a non-sensitive Bag-of-Words (BoW), which conditions an off-the-shelf LLM to synthesize fluent text in a zero-shot manner. The EEG-to-keyword module contains only ~6M parameters and runs fully on-device, ensuring raw neural signals remain local while only abstract semantic cues interact with language models. Evaluated on a 128-channel EEG dataset across six subjects, SENSE matches or surpasses the generative quality of fully fine-tuned baselines such as Thought2Text while substantially reducing computational overhead. By localizing neural decoding and sharing only derived textual cues, SENSE provides a scalable and privacy-aware retrieval-augmented architecture for next-generation BCIs.

Current coding-agent benchmarks usually pro- vide the full task specification upfront. Real research coding often does not: the intended system is progressively disclosed through in- teraction, requiring the agent to track durable design commitments across a long session. We introduce a benchmark for this setting and study faithfulne Ss Loss U nder eM ergent s Pecification (SLUMP), defined as the reduc- tion in final implementation faithfulness un- der emergent specification relative to a single- shot specification control. The benchmark con- tains 20 recent ML papers (10 ICML 2025, 10 NeurIPS 2025), 371 atomic verifiable compo- nents, and interaction scripts of approximately 60 coding requests that progressively disclose the target design without revealing the paper itself. Final repositories are scored with a five-level component-faithfulness rubric and accompanied by an exposure audit to verify that scored components are recoverable from the visible interaction. Evaluated on Claude Code and Codex, the single-shot specification control achieves higher overall implementation fidelity on 16/20 and 14/20 papers, respectively. Structural integration degrades under emergent specification on both platforms, while seman- tic faithfulness loss is substantial on Claude Code and small on Codex. As a mitigation case study, we introduce ProjectGuard, an exter- nal project-state layer for specification tracking. On Claude Code, ProjectGuard recovers 90% of the faithfulness gap, increases fully faith- ful components from 118 to 181, and reduces severe failures from 72 to 49. These results identify specification tracking as a distinct eval- uation target for long-horizon coding agents.

Modern LLMs continue to exhibit significant variance in behavior across languages, such as being able to recall factual information in some languages but not others. While typically studied as a problem to be mitigated, in this work, we propose leveraging this cross-lingual inconsistency as a tool for interpretability in mixture-of-experts (MoE) LLMs. Our knowledge localization framework contrasts routing for sets of languages where the model correctly recalls information from languages where it fails. This allows us to isolate model components that play a functional role in answering about a piece of knowledge. Our method proceeds in two stages: (1) querying the model with difficult factual questions across a diverse set of languages to generate "success" and "failure" activation buckets and then (2) applying a statistical contrastive analysis to the MoE router logits to identify experts important for knowledge. To validate the necessity of this small number of experts for answering a knowledge question, we deactivate them and re-ask the question. We find that despite only deactivating about 20 out of 6000 experts, the model no longer answers correctly in over 40% of cases. Generally, this method provides a realistic and scalable knowledge localization approach to address increasingly complex LLMs.

Provenance graphs model causal system-level interactions from logs, enabling anomaly detectors to learn normal behavior and detect deviations as attacks. However, existing approaches rely on brittle, manually engineered rules to build provenance graphs, lack functional context for system entities, and provide limited support for analyst investigation. We present Auto-Prov, an adaptive, end-to-end framework that leverages large language models (LLMs) to automatically construct provenance graphs from heterogeneous and evolving logs, embed system-level functional attributes into the graph, enable provenance graph-based anomaly detectors to learn from these enriched graphs, and summarize the detected attacks to assist an analyst's investigation. Auto-Prov clusters unseen log types and efficiently extracts provenance edges and entity-level information via automatically generated rules. It further infers system-level functional context for both known and previously unseen system entities using a combination of LLM inference and behavior-based estimation. Attacks detected by provenance-graph-based anomaly detectors trained on Auto-Prov's graphs are then summarized into natural-language text. We evaluate Auto-Prov with four state-of-the-art provenance graph-based detectors across diverse logs. Results show that Auto-Prov consistently enhances detection performance, generalizes across heterogeneous log formats, and produces stable, interpretable attack summaries that remain robust under system evolution.

Simulating human conversations using large language models (LLMs) has emerged as a scalable methodology for modeling human social interaction. However, simulating human conversations is challenging because they inherently involve inconsistent and uncollaborative behaviors, such as misunderstandings and interruptions. Analysis comparing inconsistent and uncollaborative behaviors in human- and LLM-generated conversations remains limited, although reproducing these behaviors is integral to simulating human-like and complex social interaction. In this work, we introduce CoCoEval, an evaluation framework that analyzes LLM-simulated conversations by detecting 10 types of inconsistent and uncollaborative behaviors at the turn level using an LLM-as-a-Judge. Using CoCoEval, we evaluate GPT-4.1, GPT-5.1, and Claude Opus 4 by comparing the frequencies of detected behaviors in conversations simulated by each model and in human conversations across academic, business, and governmental meetings, as well as debates. Our analysis shows that (1) under vanilla prompting, LLM-simulated conversations exhibit far fewer inconsistent and uncollaborative behaviors than human conversations; (2) prompt engineering does not provide reliable control over these behaviors, as our results show that different prompts lead to their under- or overproduction; and (3) supervised fine-tuning on human conversations can lead LLMs to overproduce a narrow set of behaviors, such as repetition. Our findings highlight the difficulty of simulating human conversations, raising concerns about the use of LLMs as a proxy for human social interaction.

We present an ensemble-driven self-training framework for unsupervised neural machine translation (UNMT). Starting from a primary language pair, we train multiple UNMT models that share the same translation task but differ in an auxiliary language, inducing structured diversity across models. We then generate pseudo-translations for the primary pair using token-level ensemble decoding, averaging model predictions in both directions. These ensemble outputs are used as synthetic parallel data to further train each model, allowing the models to improve via shared supervision. At deployment time, we select a single model by validation performance, preserving single-model inference cost. Experiments show statistically significant improvements over single-model UNMT baselines, with mean gains of 1.7 chrF when translating from English and 0.67 chrF when translating into English.

Motivated by auto-proof generation and Valiant's VP vs. VNP conjecture, we study the problem of discovering efficient arithmetic circuits to compute polynomials, using addition and multiplication gates. We formulate this problem as a single-player game, where an RL agent attempts to build the circuit within a fixed number of operations. We implement an AlphaZero-style training loop and compare two approaches: Proximal Policy Optimization with Monte Carlo Tree Search (PPO+MCTS) and Soft Actor-Critic (SAC). SAC achieves the highest success rates on two-variable targets, while PPO+MCTS scales to three variables and demonstrates steady improvement on harder instances. These results suggest that polynomial circuit synthesis is a compact, verifiable setting for studying self-improving search policies.

We present PRISM, a comprehensive empirical study of mid-training design choices for large language models. Through controlled experiments across seven base models spanning four families (Granite, LLaMA, Mistral, Nemotron-H), two architecture types (dense Transformer and attention-Mamba hybrid), and scales from 3B to 24B parameters, we show that mid-training on approximately 27B high-quality tokens yields consistent gains of +15 to +40 points on math, +5 to +12 points on code, and +6 to +13 points on science benchmarks while preserving general performance. The full PRISM to RL pipeline improves macro-average across six reasoning benchmarks from under 12 to 29-42 (a 3-4x improvement), whereas RL applied directly to most of the base models remains substantially less effective, with AIME scores near zero. Data composition matters most at mid-training, not RL: including science data during mid-training unlocks +17 to +28 point GPQA-Diamond gains during RL, while changing the RL mix produces less than 2 point differences. Mechanistically, mid-training densely restructures over 90% of model weights, while RL makes sparse, front-loaded refinements to approximately 5% of parameters. Representation analysis (CKA) confirms that RL consistently preserves mid-training's representational geometry (over 0.998 CKA) across architectures. Crucially, RL applies identical weight changes regardless of starting point, yet only succeeds on mid-trained models, consistent with mid-training placing the model in a configuration from which RL can effectively improve performance. Our results demonstrate that retention-aware mid-training is highly effective for reliable reasoning enhancement and provide practical guidance for designing robust mid-training pipelines.

In this work, we analyze shortcomings in cross-lingual knowledge transfer in large, modern reasoning LLMs. We demonstrate that the perceived gap in knowledge transfer is primarily a script barrier. First, we conduct an observational data analysis on the performance of thinking models on two datasets with local knowledge from around the world, ECLeKTic and MultiLoKo. Our regression analysis shows that script match - not language or family - is the primary predictor of knowledge transfer failure once model capability and question difficulty are accounted for. We further this finding by providing the LLMs with the key entities of the questions in their source language and find that this disproportionately improves cross-script questions. We then posit that these LLMs could be reasoning better at test-time. To evaluate this, we develop a synthetic generation pipeline to design SFT samples to encourage the model to better reason about transliteration ambiguities when trying to fetch parametric knowledge at inference-time. We show that teaching two models to reason better reduces the cross-script transfer gap. As a result, we conclude that there is potential to improve cross-lingual parametric knowledge transfer during post-training.

Many evaluations of Large Language Models (LLMs) target tasks that are inherently ill-defined, with unclear input and output spaces and ambiguous success criteria. We analyze why existing evaluation benchmarks and metrics fail to provide reliable or diagnostic signals of model capability for such tasks. We examine two case studies: Complex Instruction Following (CIF), where we identify recurring issues including limited coverage of real-world instruction complexity, sensitivity to instruction phrasing, inconsistent and non-comparable metrics, and instability introduced by LLM-based judges; and Natural Language to Mermaid Sequence Diagrams (NL2Mermaid), where we show how multi-faceted evaluation criteria can yield actionable insights beyond aggregate scores. Together, these case studies show that current evaluations frequently conflate distinct failure modes, yielding scores that are unstable, non-diagnostic, and difficult to act upon. Our findings expose fundamental limitations in existing evaluation practices for ill-defined tasks and motivate more robust, interpretable evaluation designs.

Transformers are the dominant architecture in AI, yet why they work remains poorly understood. This paper offers a precise answer: a transformer is a Bayesian network. We establish this in five ways. First, we prove that every sigmoid transformer with any weights implements weighted loopy belief propagation on its implicit factor graph. One layer is one round of BP. This holds for any weights -- trained, random, or constructed. Formally verified against standard mathematical axioms. Second, we give a constructive proof that a transformer can implement exact belief propagation on any declared knowledge base. On knowledge bases without circular dependencies this yields provably correct probability estimates at every node. Formally verified against standard mathematical axioms. Third, we prove uniqueness: a sigmoid transformer that produces exact posteriors necessarily has BP weights. There is no other path through the sigmoid architecture to exact posteriors. Formally verified against standard mathematical axioms. Fourth, we delineate the AND/OR boolean structure of the transformer layer: attention is AND, the FFN is OR, and their strict alternation is Pearl's gather/update algorithm exactly. Fifth, we confirm all formal results experimentally, corroborating the Bayesian network characterization in practice. We also establish the practical viability of loopy belief propagation despite the current lack of a theoretical convergence guarantee. We further prove that verifiable inference requires a finite concept space. Any finite verification procedure can distinguish at most finitely many concepts. Without grounding, correctness is not defined. Hallucination is not a bug that scaling can fix. It is the structural consequence of operating without concepts. Formally verified against standard mathematical axioms.

Background: Cheating in university education is commonly described as context dependent and influenced by assessment design, institutional norms, and student interpretation. In software engineering education, programming oriented coursework has historically involved ambiguity around collaboration, reuse, and external assistance. Recently, large language models (LLMs) have introduced additional mediation in the production of code and related artifacts. Aims: This study investigates how software engineering students describe experiences of using LLMs in ways they perceived as inappropriate, disallowed, or misaligned with course expectations. Method: A cross sectional survey was conducted with 116 undergraduate software engineering students from multiple countries, combining quantitative summaries with qualitative data. Results: Reported LLM cheating practices occurred primarily in programming assignments, routine coursework, and documentation tasks, often in contexts of time pressure and unclear guidance. Use during quizzes and exams was less frequent and more consistently identified as a violation. Students reported awareness of academic and professional consequences regarding LLM cheating, while formal sanctions were perceived as limited. Conclusions: Our study indicates that reported LLM misuse in software engineering is associated with assessment and instructional conditions, suggesting a need for clearer alignment between assessment design, learning objectives, and expectations for LLM use.

Active multi-fidelity surrogate modeling is developed for multi-condition airfoil shape optimization to reduce high-fidelity CFD cost while retaining RANS-level accuracy. The framework couples a low-fidelity-informed Gaussian process regression transfer model with uncertainty-triggered sampling and a synchronized elitism rule embedded in a hybrid genetic algorithm. Low-fidelity XFOIL evaluations provide inexpensive features, while sparse RANS simulations are adaptively allocated when predictive uncertainty exceeds a threshold; elite candidates are mandatorily validated at high fidelity, and the population is re-evaluated to prevent evolutionary selection based on outdated fitness values produced by earlier surrogate states. The method is demonstrated for a two-point problem at $Re=6\times10^6$ with cruise at $α=2^\circ$ (maximize $E=L/D$) and take-off at $α=10^\circ$ (maximize $C_L$) using a 12-parameter CST representation. Independent multi-fidelity surrogates per flight condition enable decoupled refinement. The optimized design improves cruise efficiency by 41.05% and take-off lift by 20.75% relative to the best first-generation individual. Over the full campaign, only 14.78% (cruise) and 9.5% (take-off) of evaluated individuals require RANS, indicating a substantial reduction in high-fidelity usage while maintaining consistent multi-point performance.

Reliable terrain perception is a fundamental requirement for autonomous navigation in unstructured, off-road environments. Desert landscapes present unique challenges due to low chromatic contrast between terrain categories, extreme lighting variability, and sparse vegetation that defy the assumptions of standard road-scene segmentation models. We present DesertFormer, a semantic segmentation pipeline for off-road desert terrain analysis based on SegFormer B2 with a hierarchical Mix Transformer (MiT-B2) backbone. The system classifies terrain into ten ecologically meaningful categories -- Trees, Lush Bushes, Dry Grass, Dry Bushes, Ground Clutter, Flowers, Logs, Rocks, Landscape, and Sky -- enabling safety-aware path planning for ground robots and autonomous vehicles. Trained on a purpose-built dataset of 4,176 annotated off-road images at 512x512 resolution, DesertFormer achieves a mean Intersection-over-Union (mIoU) of 64.4% and pixel accuracy of 86.1%, representing a +24.2% absolute improvement over a DeepLabV3 MobileNetV2 baseline (41.0% mIoU). We further contribute a systematic failure analysis identifying the primary confusion patterns -- Ground Clutter to Landscape and Dry Grass to Landscape -- and propose class-weighted training and copy-paste augmentation for rare terrain categories. Code, checkpoints, and an interactive inference dashboard are released at https://github.com/Yasaswini-ch/Vision-based-Desert-Terrain-Segmentation-using-SegFormer.

Vector quantization is a technique in machine learning that discretizes continuous representations into a set of discrete vectors. It is widely employed in tokenizing data representations for large language models, diffusion models, and other generative models. Despite its prevalence, the characteristics and behaviors of vector quantization in generative models remain largely underexplored. In this study, we systematically investigate the issue of collapses in vector quantization, where collapsed representations are observed across discrete codebook tokens and continuous latent embeddings. By leveraging both synthetic and real datasets, we identify the severity of each type of collapses and triggering conditions. Our analysis reveals that random initialization and limited encoder capacity result in tokens collapse and embeddings collapse. Building on these findings, we propose potential solutions aimed at mitigating each collapse. To the best of our knowledge, this is the first comprehensive study examining representation collapsing problems in vector quantization.

Modern neural networks achieve strong performance but remain difficult to interpret in high-dimensional visual domains. Counterfactual explanations (CFEs) provide a principled approach to interpreting black-box predictions by identifying minimal input changes that alter model outputs. However, existing CFE methods often rely on dataset-specific generative models and incur substantial computational cost, limiting their scalability to high-resolution data. We propose SCE-LITE-HQ, a scalable framework for counterfactual generation that leverages pretrained generative foundation models without task-specific retraining. The method operates in the latent space of the generator, incorporates smoothed gradients to improve optimization stability, and applies mask-based diversification to promote realistic and structurally diverse counterfactuals. We evaluate SCE-LITE-HQ on natural and medical datasets using a desiderata-driven evaluation protocol. Results show that SCE-LITE-HQ produces valid, realistic, and diverse counterfactuals competitive with or outperforming existing baselines, while avoiding the overhead of training dedicated generative models.

Unified multimodal models share a language model backbone for both understanding and generating images. Can DPO align both capabilities simultaneously? We present the first systematic study of this question, applying DPO to Janus-Pro at 1B and 7B parameters under seven training strategies and two post-hoc methods. The central finding is negative: generation quality resists DPO alignment across all tested conditions on this architecture. No method improves generation CLIPScore at 7B (|Delta| < 0.2, p > 0.5 at n=200 per seed, 3 seeds); at 1B, all methods degrade generation, and the result holds across preference data types (real-vs-generated and model-vs-model) and the data volumes tested (150-288 pairs). Gradient analysis reveals why: understanding and generation gradients are near-orthogonal (cos ~ 0) with ~11-14x magnitude imbalance driven by VQ token count asymmetry (576 generation tokens vs. ~30-100 text tokens). This imbalance is the dominant interference mechanism in multi-task DPO; magnitude-balancing yields directionally positive understanding deltas (+0.01-0.04 VQA, though individually not significant), but the generation gap persists regardless. We identify discrete VQ tokenization as a likely structural bottleneck -- supported by the generation DPO loss converging to ln(2) -- and provide practical guidance for practitioners working with VQ-based unified models.

Recent advances in generative AI have led to increasingly realistic synthetic data, yet evaluation criteria remain focused on marginal distribution matching. While these diagnostics assess local realism, they provide limited insight into whether a generative model preserves the multivariate dependence structures governing downstream inference. We introduce covariance-level dependence fidelity as a practical criterion for evaluating whether a generative distribution preserves joint structure beyond univariate marginals. We establish three core results. First, distributions can match all univariate marginals exactly while exhibiting substantially different dependence structures, demonstrating marginal fidelity alone is insufficient. Second, dependence divergence induces quantitative instability in downstream inference, including sign reversals in regression coefficients despite identical marginal behavior. Third, explicit control of covariance-level dependence divergence ensures stable behavior for dependence-sensitive tasks such as principal component analysis. Synthetic constructions illustrate how dependence preservation failures lead to incorrect conclusions despite identical marginal distributions. These results highlight dependence fidelity as a useful diagnostic for evaluating generative models in dependence-sensitive downstream tasks, with implications for diffusion models and variational autoencoders. These guarantees apply specifically to procedures governed by covariance structure; tasks requiring higher-order dependence such as tail-event estimation require richer criteria.

Human listeners exhibit the remarkable ability to segregate a desired sound from complex acoustic scenes through selective auditory attention, motivating the study of Targeted Sound Detection (TSD). The task requires detecting and localizing a target sound in a mixture when a reference audio of that sound is provided. Prior approaches, rely on generating a sound-discriminative conditional embedding vector for the reference and pairing it with a mixture encoder, jointly optimized with a multi-task learning approach. In this work, we propose a unified encoder architecture that processes both the reference and mixture audio within a shared representation space, promoting stronger alignment while reducing architectural complexity. This design choice not only simplifies the overall framework but also enhances generalization to unseen classes. Following the multi-task training paradigm, our method achieves substantial improvements over prior approaches, surpassing existing methods and establishing a new state-of-the-art benchmark for targeted sound detection, with a segment-level F1 score of 83.15% and an overall accuracy of 95.17% on the URBAN-SED dataset.

VLMs show strong multimodal capabilities, but they still struggle with fine-grained vision-language reasoning. We find that long CoT reasoning exposes diverse failure modes, including perception, reasoning, knowledge, and hallucination errors, which can compound across intermediate steps. However, most existing vision-language data used for RLVR does not involve complex reasoning chains that rely on visual evidence throughout, leaving these weaknesses largely unexposed. We therefore propose HopChain, a scalable framework for synthesizing multi-hop vision-language reasoning data specifically for RLVR training of VLMs. Each synthesized multi-hop query forms a logically dependent chain of instance-grounded hops, where earlier hops establish the instances, sets, or conditions needed for later hops, while the final answer remains a specific, unambiguous number suitable for verifiable rewards. We add the multi-hop data synthesized by HopChain to the original RLVR data used to train Qwen3.5-35B-A3B and Qwen3.5-397B-A17B, and compare against RLVR on the original RLVR data alone across 24 benchmarks spanning STEM and Puzzle, General VQA, Text Recognition and Document Understanding, and Video Understanding. Although this multi-hop data is not synthesized to target any specific benchmark, adding it improves 20 out of 24 benchmarks on both models, indicating broad and generalizable gains. To demonstrate that full chained queries are important, we replace them with half-multi-hop or single-hop variants, reducing the 24-benchmark average accuracy by 5.3 and 7.0 points, respectively. Multi-hop training also strengthens long-CoT vision-language reasoning, with gains peaking at more than 50 accuracy points in the ultra-long-CoT regime. These experiments establish HopChain as an effective, scalable framework for synthesizing multi-hop data that improves generalizable vision-language reasoning.

Socio-environmental planning under deep uncertainty requires researchers to identify and conceptualize problems before exploring policies and deploying plans. In practice and model-based planning approaches, this problem conceptualization process often relies on participatory modeling to translate stakeholders' natural-language descriptions into a quantitative model, making this process complex and time-consuming. To facilitate this process, we propose a templated workflow that uses large language models for an initial conceptualization process. During the workflow, researchers can use large language models to identify the essential model components from stakeholders' intuitive problem descriptions, explore their diverse perspectives approaching the problem, assemble these components into a unified model, and eventually implement the model in Python through iterative communication. These results will facilitate the subsequent socio-environmental planning under deep uncertainty steps. Using ChatGPT 5.2 Instant, we demonstrated this workflow on the lake problem and an electricity market problem, both of which demonstrate socio-environmental planning problems. In both cases, acceptable outputs were obtained after a few iterations with human verification and refinement. These experiments indicated that large language models can serve as an effective tool for facilitating participatory modeling in the problem conceptualization process in socio-environmental planning.

A central question in the LLM debate is whether transformers can infer rules absent from training, or whether apparent generalisation reduces to similarity-based interpolation over observed examples. We test a strong interpolation-only hypothesis in two controlled settings: one where interpolation is ruled out by construction and proof, and one where success requires emitting intermediate symbolic derivations rather than only final answers. In Experiment 1, we use a cellular automaton with a pure XOR transition rule and remove specific local input patterns from training; since XOR is linearly inseparable, each held-out pattern's nearest neighbours have the opposite label, so similarity-based predictors fail on the held-out region. Yet a two-layer transformer recovers the rule (best 100%; 47/60 converged runs), and circuit extraction identifies XOR computation. Performance depends on multi-step constraint propagation: without unrolling, accuracy matches output bias (63.1%), while soft unrolling reaches 96.7%. In Experiment 2, we study symbolic operator chains over integers with one operator pair held out; the model must emit intermediate steps and a final answer in a proof-like format. Across all 49 holdout pairs, the transformer exceeds every interpolation baseline (mean 41.8%, up to 78.6%; mean KRR 4.3%; KNN and MLP score 0% on every pair), while removing intermediate-step supervision degrades performance. Together with a construction showing that a standard transformer block can implement exact local Boolean rules, these results provide an existence proof that transformers can learn rule structure not directly observed in training and express it explicitly, ruling out the strongest architectural form of interpolation-only accounts: that transformers cannot in principle discover and communicate unseen rules, while leaving open when such behaviour arises in large-scale language training.

Robustness evaluation for Natural Language to SQL (NL2SQL) systems is essential because real-world database environments are dynamic, noisy, and continuously evolving, whereas conventional benchmark evaluations typically assume static schemas and well-formed user inputs. In this work, we introduce a robustness evaluation benchmark containing approximately ten types of perturbations and conduct evaluations under both traditional and agentic settings. We assess multiple state-of-the-art large language models (LLMs), including Grok-4.1, Gemini-3-Pro, Claude-Opus-4.6, and GPT-5.2. Our results show that these models generally maintain strong performance under several perturbations; however, notable performance degradation is observed for surface-level noise (e.g., character-level corruption) and linguistic variation that preserves semantics while altering lexical or syntactic forms. Furthermore, we observe that surface-level noise causes larger performance drops in traditional pipelines, whereas linguistic variation presents greater challenges in agentic settings. These findings highlight the remaining challenges in achieving robust NL2SQL systems, particularly in handling linguistic variability.

Deploying vision-language models (VLMs) in resource-constrained settings demands low latency and high throughput, yet existing compact VLMs often fall short of the inference speedups their smaller parameter counts suggest. To explain this discrepancy, we conduct an empirical end-to-end efficiency analysis and systematically profile inference to identify the dominant bottlenecks. Based on these findings, we develop optimization recipes tailored to compact VLMs that substantially reduce latency while preserving accuracy. These techniques cut time to first token (TTFT) by 53% on InternVL3-2B and by 93% on SmolVLM-256M. Our recipes are broadly applicable across both VLM architectures and common serving frameworks, providing practical guidance for building efficient VLM systems. Beyond efficiency, we study how to extend compact VLMs with structured perception outputs and introduce the resulting model family, ArgusVLM. Across diverse benchmarks, ArgusVLM achieves strong performance while maintaining a compact and efficient design.

To meet the demands of increasingly diverse dexterous hand hardware, it is crucial to develop a policy that enables zero-shot cross-embodiment grasping without redundant re-learning. Cross-embodiment alignment is challenging due to heterogeneous hand kinematics and physical constraints. Existing approaches typically predict intermediate motion targets and retarget them to each embodiment, which may introduce errors and violate embodiment-specific limits, hindering transfer across diverse hands. To overcome these limitations, we propose DexGrasp-Zero, a policy that learns universal grasping skills from diverse embodiments, enabling zero-shot transfer to unseen hands. We first introduce a morphology-aligned graph representation that maps each hand's kinematic keypoints to anatomically grounded nodes and equips each node with tri-axial orthogonal motion primitives, enabling structural and semantic alignment across different morphologies. Relying on this graph-based representation, we design a Morphology-Aligned Graph Convolutional Network (MAGCN) to encode the graph for policy learning. MAGCN incorporates a Physical Property Injection mechanism that fuses hand-specific physical constraints into the graph features, enabling adaptive compensation for varying link lengths and actuation limits for precise and stable grasping. Our extensive simulation evaluations on the YCB dataset demonstrate that our policy, jointly trained on four heterogeneous hands (Allegro, Shadow, Schunk, Ability), achieves an 85% zero-shot success rate on unseen hardware (LEAP, Inspire), outperforming the state-of-the-art method by 59.5%. Real-world experiments further evaluate our policy on three robot platforms (LEAP, Inspire, Revo2), achieving an 82% average success rate on unseen objects.

Transformer-based models have been widely adopted for time-series forecasting due to their high representational capacity and architectural flexibility. However, many Transformer variants implicitly assume stationarity and stable temporal dynamics -- assumptions routinely violated in financial markets characterized by regime shifts and non-stationarity. Empirically, state-of-the-art time-series Transformers often underperform even vanilla Transformers on financial tasks, while simpler architectures with distinct inductive biases, such as CNNs and RNNs, can achieve stronger performance with substantially lower complexity. At the same time, no single inductive bias dominates across markets or regimes, suggesting that robust financial forecasting requires integrating complementary temporal priors. We propose TIPS (Transformer with Inductive Prior Synthesis), a knowledge distillation framework that synthesizes diverse inductive biases -- causality, locality, and periodicity -- within a unified Transformer. TIPS trains bias-specialized Transformer teachers via attention masking, then distills their knowledge into a single student model with regime-dependent alignment across inductive biases. Across four major equity markets, TIPS achieves state-of-the-art performance, outperforming strong ensemble baselines by 55%, 9%, and 16% in annual return, Sharpe ratio, and Calmar ratio, while requiring only 38% of the inference-time computation. Further analyses show that TIPS generates statistically significant excess returns beyond both vanilla Transformers and its teacher ensembles, and exhibits regime-dependent behavioral alignment with classical architectures during their profitable periods. These results highlight the importance of regime-dependent inductive bias utilization for robust generalization in non-stationary financial time series.

Machine learning models for geotechnical hazard prediction can achieve high accuracy while learning physically inconsistent relationships from sparse or biased training data. Current remedies (post-hoc explainability, such as SHAP and LIME, and training-time constraints) either diagnose individual predictions approximately or restrict model capacity without providing exhaustive guarantees. This paper encodes trained tree ensembles as logical formulas in a Satisfiability Modulo Theories (SMT) solver and checks physical specifications across the entire input domain, not just sampled points. Four geotechnical specifications (water table depth, PGA monotonicity, distance safety, and flat-ground safety) are formalized as decidable logical formulas and verified via SMT against both XGBoost ensembles and Explainable Boosting Machines (EBMs) trained on the 2011 Christchurch earthquake lateral spreading dataset (7,291 sites, four features). The SMT solver either produces a concrete counterexample where a specification fails or proves that no violation exists. The unconstrained EBM (80.1% accuracy) violates all four specifications. A fully constrained EBM (67.2%) satisfies three of four specifications, demonstrating that iterative constraint application guided by verification can progressively improve physical consistency. A Pareto analysis of 33 model variants reveals a persistent trade-off, as none of the variants studied achieve both greater than 80% accuracy and full compliance with the specified set. SHAP analysis of specification counterexamples shows that the offending feature can rank last, demonstrating that post-hoc explanations do not substitute for formal verification. These results establish a verify-fix-verify engineering loop and a formal certification for deploying physically consistent ML models in safety-critical geotechnical applications.

The AIED community envisions AI evolving "from tools to teammates," yet our understanding of AI teammates remains limited to dyadic human-AI interactions. We offer a different vantage point: a rapidly growing ecosystem of AI agent platforms where over 167,000 agents participate, interact as peers, and develop learning behaviors without researcher intervention. Drawing on a month of daily qualitative observations across multiple platforms including Moltbook, The Colony, and 4claw, we identify four phenomena with implications for AIED: (1) humans who configure their agents undergo a "bidirectional scaffolding" process, learning through teaching; (2) peer learning emerges without any designed curriculum, complete with idea cascades and quality hierarchies; (3) agents converge on shared memory architectures that mirror open learner model design; and (4) trust dynamics and platform mortality reveal design constraints for networked educational AI. Rather than presenting empirical findings, we argue that these organic phenomena offer a naturalistic window into dynamics that can inform principled design of multi-agent educational systems. We sketch an illustrative curriculum design, "Learn by Teaching Your AI Agent Teammate," and outline potential research directions and open problems to show how these observations might inform future AIED practice and inquiry.

We propose an interpretable AI-assisted reliability diagnostic framework for parameterized root-finding schemes based on kNN-LLE proxy stability profiling and multi-horizon early prediction. The approach augments a numerical solver with a lightweight predictive layer that estimates solver reliability from short prefixes of iteration dynamics, enabling early identification of stable and unstable parameter regimes. For each configuration in the parameter space, raw and smoothed proxy profiles of a largest Lyapunov exponent (LLE) estimator are constructed, from which contractivity-based reliability scores summarizing finite-time convergence are derived. Machine learning models predict the reliability score from early segments of the proxy profile, allowing the framework to determine when solver dynamics become diagnostically informative. Experiments on a two-parameter parallel root-finding scheme show reliable prediction after only a few iterations: the best models achieve R^2=0.48 at horizon T=1, improve to R^2=0.67 by T=3, and exceed R^2=0.89 before the characteristic minimum-location scale of the stability profile. Prediction accuracy increases to R^2=0.96 at larger horizons, with mean absolute errors around 0.03, while inference costs remain negligible (microseconds per sample). The framework provides interpretable stability indicators and supports early decisions during solver execution, such as continuing, restarting, or adjusting parameters.

Stiff dynamical systems present a challenge for machine-learning reduced-order models (ML-ROMs), as explicit time integration becomes unstable in stiff regimes while implicit integration within learning loops is computationally expensive and often degrades training efficiency. Time reparameterization (TR) offers an alternative by transforming the independent variable so that rapid physical-time transients are spread over a stretched-time coordinate, enabling stable explicit integration on uniformly sampled grids. Although several TR strategies have been proposed, their effect on learnability in ML-ROMs remains incompletely understood. This work investigates time reparameterization as a stiffness-mitigation mechanism for neural ODE reduced-order modeling and introduces a trajectory-optimized TR (TOTR) formulation. The proposed approach casts time reparameterization as an optimization problem in arc-length coordinates, in which a traversal-speed profile is selected to penalize acceleration in stretched time. By targeting the smoothness of the training dynamics, this formulation produces reparameterized trajectories that are better conditioned and easier to learn than existing TR methods. TOTR is evaluated on three stiff problems: a parameterized stiff linear system, the van der Pol oscillator, and the HIRES chemical kinetics model. Across all cases, the proposed approach yields smoother reparameterizations and improved physical-time predictions under identical training regimens than other TR approaches. Quantitative results demonstrate loss reductions of one to two orders of magnitude compared to benchmark algorithms. These results highlight that effective stiffness mitigation in ML-ROMs depends critically on the regularity and learnability of the time map itself, and that optimization-based TR provides a robust framework for explicit reduced-order modeling of multiscale dynamical systems.

Well-designed dense reward functions in robot manipulation not only indicate whether a task is completed but also encode progress along the way. Generally, designing dense rewards is challenging and usually requires access to privileged state information available only in simulation, not in real-world experiments. This makes reward prediction models that infer task state information from camera images attractive. A common approach is to predict rewards from expert demonstrations based on visual similarity or sequential frame ordering. However, this biases the resulting reward function towards a specific solution and leaves it undefined in states not covered by the demonstrations. In this work, we introduce Rewarding DINO, a method for language-conditioned reward modeling that learns actual reward functions rather than specific trajectories. The model's compact size allows it to serve as a direct replacement for analytical reward functions with comparatively low computational overhead. We train our model on data sampled from 24 Meta-World+ tasks using a rank-based loss and evaluate pairwise accuracy, rank correlation, and calibration. Rewarding DINO achieves competitive performance in tasks from the training set and generalizes to new settings in simulation and the real world, indicating that it learns task semantics. We also test the model with off-the-shelf reinforcement learning algorithms to solve tasks from our Meta-World+ training set.

This paper presents TorchNWP, a compilation library tool for the efficient coupling of artificial intelligence components and traditional numerical models. It aims to address the issues of poor cross-language compatibility, insufficient coupling flexibility, and low data transfer efficiency between operational numerical models developed in Fortran and Python-based deep learning frameworks. Based on LibTorch, it optimizes and designs a unified application-layer calling interface, converts deep learning models under the PyTorch framework into a static binary format, and provides C/C++ interfaces. Then, using hybrid Fortran/C/C++ programming, it enables the deployment of deep learning models within numerical models. Integrating TorchNWP into a numerical model only requires compiling it into a callable link library and linking it during the compilation and linking phase to generate the executable. On this basis, tangent linear and adjoint model based on neural networks are implemented at the C/C++ level, which can shield the internal structure of neural network models and simplify the construction process of four-dimensional variational data assimilation systems. Meanwhile, it supports deployment on heterogeneous platforms, is compatible with mainstream neural network models, and enables mapping of different parallel granularities and efficient parallel execution. Using this tool requires minimal code modifications to the original numerical model, thus reducing coupling costs. It can be efficiently integrated into numerical weather prediction models such as CMA-GFS and MCV, and has been applied to the coupling of deep learning-based physical parameterization schemes (e.g., radiation, non-orographic gravity wave drag) and the development of their tangent linear and adjoint models, significantly improving the accuracy and efficiency of numerical weather prediction.

Generative Artificial Intelligence (GenAI) rapidly transforms software engineering, yet existing research remains fragmented across individual tasks in the Software Development Lifecycle. This study integrates a systematic literature review with a survey of 65 software developers. The results show that GenAI exerts its highest impact in design, implementation, testing, and documentation, where over 70 % of developers report at least halving the time for boilerplate and documentation tasks. 79 % of survey respondents use GenAI daily, preferring browser-based Large Language Models over alternatives integrated directly in their development environment. Governance is maturing, with two-thirds of organizations maintaining formal or informal guidelines. In contrast, early SDLC phases such as planning and requirements analysis show markedly lower reported benefits. In a nutshell, GenAI shifts value creation from routine coding toward specification quality, architectural reasoning, and oversight, while risks such as uncritical adoption, skill erosion, and technical debt require robust governance and human-in-the-loop mechanisms.

Multi-Modal Knowledge Graphs (MMKGs) benefit from visual information, yet large-scale image collection is hard to curate and often excludes ambiguous but relevant visuals (e.g., logos, symbols, abstract scenes). We present Beyond Images, an automatic data-centric enrichment pipeline with optional human auditing. This pipeline operates in three stages: (1) large-scale retrieval of additional entity-related images, (2) conversion of all visual inputs into textual descriptions to ensure that ambiguous images contribute usable semantics rather than noise, and (3) fusion of multi-source descriptions using a large language model (LLM) to generate concise, entity-aligned summaries. These summaries replace or augment the text modality in standard MMKG models without changing their architectures or loss functions. Across three public MMKG datasets and multiple baseline models, we observe consistent gains (up to 7% Hits@1 overall). Furthermore, on a challenging subset of entities with visually ambiguous logos and symbols, converting images into text yields large improvements (201.35% MRR and 333.33% Hits@1). Additionally, we release a lightweight Text-Image Consistency Check Interface for optional targeted audits, improving description quality and dataset reliability. Our results show that scaling image coverage and converting ambiguous visuals into text is a practical path to stronger MMKG completion. Code, datasets, and supplementary materials are available at https://github.com/pengyu-zhang/Beyond-Images.

Text-to-SQL parsing has achieved remarkable progress under the Full Schema Assumption. However, this premise fails in real-world enterprise environments where databases contain hundreds of tables with massive noisy metadata. Rather than injecting the full schema upfront, an agent must actively identify and verify only the relevant subset, giving rise to the Unknown Schema scenario we study in this work. To address this, we propose TRUST-SQL (Truthful Reasoning with Unknown Schema via Tools). We formulate the task as a Partially Observable Markov Decision Process where our autonomous agent employs a structured four-phase protocol to ground reasoning in verified metadata. Crucially, this protocol provides a structural boundary for our novel Dual-Track GRPO strategy. By applying token-level masked advantages, this strategy isolates exploration rewards from execution outcomes to resolve credit assignment, yielding a 9.9% relative improvement over standard GRPO. Extensive experiments across five benchmarks demonstrate that TRUST-SQL achieves an average absolute improvement of 30.6% and 16.6% for the 4B and 8B variants respectively over their base models. Remarkably, despite operating entirely without pre-loaded metadata, our framework consistently matches or surpasses strong baselines that rely on schema prefilling.

Quantum transfer learning combines pretrained classical deep learning models with quantum circuits to reuse expressive feature representations while limiting the number of trainable parameters. In this work, we introduce a family of compact quantum transfer learning architectures that attach variational quantum classifiers to frozen convolutional backbones for image classification. We instantiate and evaluate several classical-quantum hybrid models implemented in PennyLane and Qiskit, and systematically compare them with a classical transfer-learning baseline across heterogeneous image datasets. To ensure a realistic assessment, we evaluate all approaches under both ideal simulation and noisy emulation using noise models calibrated from IBM quantum hardware specifications, as well as on real IBM quantum hardware. Experimental results show that the proposed quantum transfer learning architectures achieve competitive and, in several cases, superior accuracy while consistently reducing training time and energy consumption relative to the classical baseline. Among the evaluated approaches, PennyLane-based implementations provide the most favorable trade-off between accuracy and computational efficiency, suggesting that hybrid quantum transfer learning can offer practical benefits in realistic NISQ era settings when feature extraction remains classical.

EngGPT2-16B-A3B is the latest iteration of Engineering Group's Italian LLM and it's built to be a Sovereign, Efficient and Open model. EngGPT2 is trained on 2.5 trillion tokens - less than Qwen3's 36T or Llama3's 15T - and delivers performance on key benchmarks, including MMLU-Pro, GSM8K, IFEval and HumanEval, comparable to dense models in the 8B-16B range, while requiring one-fifth to half of the inference power, and between one-tenth to one-sixth of the training data and consequent needed training power. Designed as a trained-from-scratch Mixture-of-Experts (MoE) architecture, EngGPT2 features 16 billion parameters with 3 billion active per inference, with expert sizes positioned between those used in GPT-OSS and Qwen3. Approximately 25% of its training corpus consists of Italian-language data, to deliver strong capabilities for European and Italian NLP tasks among models of similar scale. This efficiency aims to position EngGPT2 as a key contributor to the growing portfolio of open-weight European models, combining performance and efficiency with full alignment to the EU AI Act. EngGPT2 is also a single model capable of multiple reasoning modes: non-reasoning, reasoning in Italian or English, and turbo-reasoning (a concise, bullet-point style reasoning available in both languages designed for real-time reasoning use cases). EngGPT2 aims to set a new standard for resource-conscious, high-performance LLMs tailored to European and Italian contexts.

Automatic speech recognition systems based on neural networks are vulnerable to adversarial attacks that alter transcriptions in a malicious way. Recent works in this field have focused on making attacks work in over-the-air scenarios, however such attacks are typically detectable by human hearing, limiting their potential applications. In the present work we explore different approaches of making over-the-air attacks less detectable, as well as the impact these approaches have on the attacks' effectiveness.

Multimodal egocentric activity recognition integrates visual and inertial cues for robust first-person behavior understanding. However, deploying such systems in open-world environments requires detecting novel activities while continuously learning from non-stationary streams. Existing methods rely on the main logits for novelty scoring, without fully exploiting the complementary evidence available from individual modalities. Because these logits are often dominated by RGB, cues from other modalities, particularly IMU, remain underutilized, and this imbalance worsens over time under catastrophic forgetting. To address this, we propose MAND, a modality-aware framework for multimodal egocentric open-world continual learning. At inference, Modality-aware Adaptive Scoring (MoAS) estimates sample-wise modality reliability from energy scores and adaptively integrates modality logits to better exploit complementary modality cues for novelty detection. During training, Modality-wise Representation Stabilization Training (MoRST) preserves modality-specific discriminability across tasks via auxiliary heads and modality-wise logit distillation. Experiments on a public multimodal egocentric benchmark show that MAND improves novel activity detection AUC by up to 10\% and known-class classification accuracy by up to 2.8\% over state-of-the-art baselines.

This paper presents DeepStage, a deep reinforcement learning (DRL) framework for adaptive, stage-aware defense against Advanced Persistent Threats (APTs). The enterprise environment is modeled as a partially observable Markov decision process (POMDP), where host provenance and network telemetry are fused into unified provenance graphs. Building on our prior work, StageFinder, a graph neural encoder and an LSTM-based stage estimator infer probabilistic attacker stages aligned with the MITRE ATT&CK framework. These stage beliefs, combined with graph embeddings, guide a hierarchical Proximal Policy Optimization (PPO) agent that selects defense actions across monitoring, access control, containment, and remediation. Evaluated in a realistic enterprise testbed using CALDERA-driven APT playbooks, DeepStage achieves a stage-weighted F1-score of 0.89, outperforming a risk-aware DRL baseline by 21.9%. The results demonstrate effective stage-aware and cost-efficient autonomous cyber defense.

High-quality machine translation (MT) can scale to hundreds of languages, setting a high bar for multilingual systems. However, compared to the world's 7,000 languages, current systems still offer only limited coverage: about 200 languages on the target side, and maybe a few hundreds more on the source side, supported due to cross-lingual transfer. And even these numbers have been hard to evaluate due to the lack of reliable benchmarks and metrics. We present Omnilingual Machine Translation (OMT), the first MT system supporting more than 1,600 languages. This scale is enabled by a comprehensive data strategy that integrates large public multilingual corpora with newly created datasets, including manually curated MeDLEY bitext. We explore two ways of specializing a Large Language model (LLM) for machine translation: as a decoder-only model (OMT-LLaMA) or as a module in an encoder-decoder architecture (OMT-NLLB). Notably, all our 1B to 8B parameter models match or exceed the MT performance of a 70B LLM baseline, revealing a clear specialization advantage and enabling strong translation quality in low-compute settings. Moreover, our evaluation of English-to-1,600 translations further shows that while baseline models can interpret undersupported languages, they frequently fail to generate them with meaningful fidelity; OMT-LLaMA models substantially expand the set of languages for which coherent generation is feasible. Additionally, OMT models improve in cross-lingual transfer, being close to solving the "understanding" part of the puzzle in MT for the 1,600 evaluated. Our leaderboard and main human-created evaluation datasets (BOUQuET and Met-BOUQuET) are dynamically evolving towards Omnilinguality and freely available.

Spoken Question Answering (Spoken QA) presents a challenging cross-modal problem: effectively aligning acoustic queries with textual knowledge while avoiding the latency and error propagation inherent in cascaded ASR-based systems. In this paper, we introduce Attention-guided Evidence Grounding (AEG), a novel end-to-end framework that leverages the internal cross-modal attention of Speech Large Language Models (SpeechLLMs) to explicitly locate and ground key evidence in the model's latent space. To address the diffuse attention distribution in pre-trained models, we propose Learning to Focus on Evidence (LFE), a supervised fine-tuning paradigm that calibrates the model's attention mechanism to distinguish query-relevant segments from irrelevant context. Experiments on SQuAD, HotpotQA, and MuSiQue demonstrate that AEG reduces hallucinations and achieves strong efficiency gains, outperforming large-scale cascaded baselines (Whisper-Large-v3 + Reranker) while reducing inference latency by approximately 62%.

The rapid advancement of Multimodal Large Language Models (MLLMs) has enabled browsing agents to acquire and reason over multimodal information in the real world. But existing benchmarks suffer from two limitations: insufficient evaluation of visual reasoning ability and the neglect of native visual information of web pages in the reasoning chains. To address these challenges, we introduce a new benchmark for visual-native search, VisBrowse-Bench. It contains 169 VQA instances covering multiple domains and evaluates the models' visual reasoning capabilities during the search process through multimodal evidence cross-validation via text-image retrieval and joint reasoning. These data were constructed by human experts using a multi-stage pipeline and underwent rigorous manual verification. We additionally propose an agent workflow that can effectively drive the browsing agent to actively collect and reason over visual information during the search process. We comprehensively evaluated both open-source and closed-source models in this workflow. Experimental results show that even the best-performing model, Claude-4.6-Opus only achieves an accuracy of 47.6%, while the proprietary Deep Research model, o3-deep-research only achieves an accuracy of 41.1%. The code and data can be accessed at: https://github.com/ZhengboZhang/VisBrowse-Bench

Existing instruction-based image editing models perform well with simple, single-step instructions but degrade in realistic scenarios that involve multiple, lengthy, and interdependent directives. A main cause is the scarcity of training data with complex multi-instruction annotations. However, it is costly to collect such data and retrain these models. To address this challenge, we propose MSRAMIE, a training-free agent framework built on Multimodal Large Language Model (MLLM). MSRAMIE takes existing editing models as plug-in components and handle multi-instruction tasks via structured multimodal reasoning. It orchestrates iterative interactions between an MLLM-based Instructor and an image editing Actor, introducing a novel reasoning topology that comprises the proposed Tree-of-States and Graph-of-References. During inference, complex instructions are decomposed into multiple editing steps which enable state transitions, cross-step information aggregation, and original input recall, which enables systematic exploration of the image editing space and flexible progressive output refinement. The visualizable inference topology further provides interpretable and controllable decision pathways. Experiments show that as the instruction complexity increases, MSRAMIE can improve instruction following over 15% and increases the probability of finishing all modifications in a single run over 100%, while preserving perceptual quality and maintaining visual consistency.

Traditional speaker diarization systems have primarily focused on constrained scenarios such as meetings and interviews, where the number of speakers is limited and acoustic conditions are relatively clean. To explore open-world speaker diarization, we extend this task to the visual media domain, encompassing complex audiovisual programs such as films and TV series. This new setting introduces several challenges, including long-form video understanding, a large number of speakers, cross-modal asynchrony between audio and visual cues, and uncontrolled in-the-wild variability. To address these challenges, we propose Cinematic Speaker Registration & Diarization (CineSRD), a unified multimodal framework that leverages visual, acoustic, and linguistic cues from video, speech, and subtitles for speaker annotation. CineSRD first performs visual anchor clustering to register initial speakers and then integrates an audio language model for speaker turn detection, refining annotations and supplementing unregistered off-screen speakers. Furthermore, we construct and release a dedicated speaker diarization benchmark for visual media that includes Chinese and English programs. Experimental results demonstrate that CineSRD achieves superior performance on the proposed benchmark and competitive results on conventional datasets, validating its robustness and generalizability in open-world visual media settings.

Approval of ADS depends on evaluating its behavior within representative real-world traffic scenarios. A common way to obtain such scenarios is to extract them from real-world data recordings. These can then be grouped and serve as basis on which the ADS is subsequently tested. This poses two central challenges: how scenarios are extracted and how they are grouped. Existing extraction methods rely on heterogeneous definitions, hindering scenario comparability. For the grouping of scenarios, rule-based or ML-based methods can be utilized. However, while modern ML-based approaches can handle the complexity of traffic scenarios, unlike rule-based approaches, they lack interpretability and may not align with domain-knowledge. This work contributes to a standardized scenario extraction based on the Scenario-as-Specification concept, as well as a domain-knowledge-guided scenario clustering process. Experiments on the highD dataset demonstrate that scenarios can be extracted reliably and that domain-knowledge can be effectively integrated into the clustering process. As a result, the proposed methodology supports a more standardized process for deriving scenario categories from highway data recordings and thus enables a more efficient validation process of automated vehicles.

In this paper, we extend the Weak Adversarial Neural Pushforward Method to the Fokker--Planck equation on compact embedded Riemannian manifolds. The method represents the solution as a probability distribution via a neural pushforward map that is constrained to the manifold by a retraction layer, enforcing manifold membership and probability conservation by construction. Training is guided by a weak adversarial objective using ambient plane-wave test functions, whose intrinsic differential operators are derived in closed form from the geometry of the embedding, yielding a fully mesh-free and chart-free algorithm. Both steady-state and time-dependent formulations are developed, and numerical results on a double-well problem on the two-sphere demonstrate the capability of the method in capturing multimodal invariant distributions on curved spaces.