The Inference Report

Large language model (LLM)-based coding agents achieve impressive results on controlled benchmarks yet routinely produce pull requests that real maintainers reject. The root cause is not functional incorrectness but a lack of organicity: generated code ignores project-specific conventions, duplicates functionality already provided by internal APIs, and violates implicit architectural constraints accumulated over years of development. Simply exposing an agent to the latest repository snapshot is not enough: the snapshot reveals the final state of the codebase, but not the repository-specific change patterns by which that state was reached. We introduce Learning to Commit, a framework that closes this gap through Online Repository Memory. Given a repository with a strict chronological split, the agent performs supervised contrastive reflection on earlier commits: it blindly attempts to resolve each historical issue, compares its prediction against the oracle diff, and distils the gap into a continuously growing set of skills-reusable patterns capturing coding style, internal API usage, and architectural invariants. When a new PR description arrives, the agent conditions its generation on these accumulated skills, producing changes grounded in the project's own evolution rather than generic pretraining priors. Evaluation is conducted on genuinely future, merged pull requests that could not have been seen during the skill-building phase, and spans multiple dimensions including functional correctness, code-style consistency, internal API reuse rate, and modified-region plausibility. Experiments on an expert-maintained repository with rich commit history show that Online Repository Memory effectively improves organicity scores on held-out future tasks.

Weight tying, i.e. sharing parameters between input and output embedding matrices, is common practice in language model design, yet its impact on the learned embedding space remains poorly understood. In this paper, we show that tied embedding matrices align more closely with output (unembedding) matrices than with input embeddings of comparable untied models, indicating that the shared matrix is shaped primarily for output prediction rather than input representation. This unembedding bias arises because output gradients dominate early in training. Using tuned lens analysis, we show this negatively affects early-layer computations, which contribute less effectively to the residual stream. Scaling input gradients during training reduces this bias, providing causal evidence for the role of gradient imbalance. This is mechanistic evidence that weight tying optimizes the embedding matrix for output prediction, compromising its role in input representation. These results help explain why weight tying can harm performance at scale and have implications for training smaller LLMs, where the embedding matrix contributes substantially to total parameter count.

Lack of accessible and dexterous robot hardware has been a significant bottleneck to achieving human-level dexterity in robots. Last year, we released Ruka, a fully open-sourced, tendon-driven humanoid hand with 11 degrees of freedom - 2 per finger and 3 at the thumb - buildable for under $1,300. It was one of the first fully open-sourced humanoid hands, and introduced a novel data-driven approach to finger control that captures tendon dynamics within the control system. Despite these contributions, Ruka lacked two degrees of freedom essential for closely imitating human behavior: wrist mobility and finger adduction/abduction. In this paper, we introduce Ruka-v2: a fully open-sourced, tendon-driven humanoid hand featuring a decoupled 2-DOF parallel wrist and abduction/adduction at the fingers. The parallel wrist adds smooth, independent flexion/extension and radial/ulnar deviation, enabling manipulation in confined environments such as cabinets. Abduction enables motions such as grasping thin objects, in-hand rotation, and calligraphy. We present the design of Ruka-v2 and evaluate it against Ruka through user studies on teleoperated tasks, finding a 51.3% reduction in completion time and a 21.2% increase in success rate. We further demonstrate its full range of applications for robot learning: bimanual and single-arm teleoperation across 13 dexterous tasks, and autonomous policy learning on 3 tasks. All 3D print files, assembly instructions, controller software, and videos are available at https://ruka-hand-v2.github.io/ .

Equivariance is a fundamental property in computer vision models, yet strict equivariance is rarely satisfied in real-world data, which can limit a model's performance. Controlling the degree of equivariance is therefore desirable. We propose a general framework for constructing soft equivariant models by projecting the model weights into a designed subspace. The method applies to any pre-trained architecture and provides theoretical bounds on the induced equivariance error. Empirically, we demonstrate the effectiveness of our method on multiple pre-trained backbones, including ViT and ResNet, across image classification, semantic segmentation, and human-trajectory prediction tasks. Notably, our approach improves the performance while simultaneously reducing equivariance error on the competitive ImageNet benchmark.

We introduce PerceptionComp, a manually annotated benchmark for complex, long-horizon, perception-centric video reasoning. PerceptionComp is designed so that no single moment is sufficient: answering each question requires multiple temporally separated pieces of visual evidence and compositional constraints under conjunctive and sequential logic, spanning perceptual subtasks such as objects, attributes, relations, locations, actions, and events, and requiring skills including semantic recognition, visual correspondence, temporal reasoning, and spatial reasoning. The benchmark contains 1,114 highly complex questions on 279 videos from diverse domains including city walk tours, indoor villa tours, video games, and extreme outdoor sports, with 100% manual annotation. Human studies show that PerceptionComp requires substantial test-time thinking and repeated perception steps: participants take much longer than on prior benchmarks, and accuracy drops to near chance (18.97%) when rewatching is disallowed. State-of-the-art MLLMs also perform substantially worse on PerceptionComp than on existing benchmarks: the best model in our evaluation, Gemini-3-Flash, reaches only 45.96% accuracy in the five-choice setting, while open-source models remain below 40%. These results suggest that perception-centric long-horizon video reasoning remains a major bottleneck, and we hope PerceptionComp will help drive progress in perceptual reasoning.

Recent advances in large language models have improved the capabilities of coding agents, yet systematic evaluation of complex, end-to-end website development remains limited. To address this gap, we introduce Vision2Web, a hierarchical benchmark for visual website development, spanning from static UI-to-code generation, interactive multi-page frontend reproduction, to long-horizon full-stack website development. The benchmark is constructed from real-world websites and comprises a total of 193 tasks across 16 categories, with 918 prototype images and 1,255 test cases. To support flexible, thorough and reliable evaluation, we propose workflow-based agent verification paradigm based on two complementary components: a GUI agent verifier and a VLM-based judge. We evaluate multiple visual language models instantiated under different coding-agent frameworks, revealing substantial performance gaps at all task levels, with state-of-the-art models still struggling on full-stack development.

We study the stochastic multi-armed bandit (MAB) problem where an underlying network structure enables side-observations across related actions. We use a bipartite graph to link actions to a set of unknowns, such that selecting an action reveals observations for all the unknowns it is connected to. While previous works rely on the assumption that all actions are permanently accessible, we investigate the more practical setting of stochastic availability, where the set of feasible actions (the "activation set") varies dynamically in each round. This framework models real-world systems with both structural dependencies and volatility, such as social networks where users provide side-information about their peers' preferences, yet are not always online to be queried. To address this challenge, we propose UCB-LP-A, a novel policy that leverages a Linear Programming (LP) approach to optimize exploration-exploitation trade-offs under stochastic availability. Unlike standard network bandit algorithms that assume constant access, UCB-LP-A computes an optimal sampling distribution over the realizable activation sets, ensuring that the necessary observations are gathered using only the currently active arms. We derive a theoretical upper bound on the regret of our policy, characterizing the impact of both the network structure and the activation probabilities. Finally, we demonstrate through numerical simulations that UCB-LP-A significantly outperforms existing heuristics that ignore either the side-information or the availability constraints.

We present Automatic Laplace Collapsed Sampling (ALCS), a general framework for marginalising latent parameters in Bayesian models using automatic differentiation, which we combine with nested sampling to explore the hyperparameter space in a robust and efficient manner. At each nested sampling likelihood evaluation, ALCS collapses the high-dimensional latent variables $z$ to a scalar contribution via maximum a posteriori (MAP) optimisation and a Laplace approximation, both computed using autodiff. This reduces the effective dimension from $d_θ+ d_z$ to just $d_θ$, making Bayesian evidence computation tractable for high-dimensional settings without hand-derived gradients or Hessians, and with minimal model-specific engineering. The MAP optimisation and Hessian evaluation are parallelised across live points on GPU-hardware, making the method practical at scale. We also show that automatic differentiation enables local approximations beyond Laplace to parametric families such as the Student-$t$, which improves evidence estimates for heavy-tailed latents. We validate ALCS on a suite of benchmarks spanning hierarchical, time-series, and discrete-likelihood models and establish where the Gaussian approximation holds. This enables a post-hoc ESS diagnostic that localises failures across hyperparameter space without expensive joint sampling.

Empowered by large-scale training, vision-language models (VLMs) achieve strong image and video understanding, yet their ability to perform spatial reasoning in both static scenes and dynamic videos remains limited. Recent advances try to handle this limitation by injecting geometry tokens from pretrained 3D foundation models into VLMs. Nevertheless, we observe that naive token fusion followed by standard fine-tuning in this line of work often leaves such geometric cues underutilized for spatial reasoning, as VLMs tend to rely heavily on 2D visual cues. In this paper, we propose GeoSR, a framework designed to make geometry matter by encouraging VLMs to actively reason with geometry tokens. GeoSR introduces two key components: (1) Geometry-Unleashing Masking, which strategically masks portions of 2D vision tokens during training to weaken non-geometric shortcuts and force the model to consult geometry tokens for spatial reasoning; and (2) Geometry-Guided Fusion, a gated routing mechanism that adaptively amplifies geometry token contributions in regions where geometric evidence is critical. Together, these designs unleash the potential of geometry tokens for spatial reasoning tasks. Extensive experiments on both static and dynamic spatial reasoning benchmarks demonstrate that GeoSR consistently outperforms prior methods and establishes new state-of-the-art performance by effectively leveraging geometric information. The project page is available at https://suhzhang.github.io/GeoSR/.

Malware continues to be a predominant operational risk for organizations, especially when obfuscation techniques are used to evade detection. Despite the ongoing efforts in the development of Machine Learning (ML) detection approaches, there is still a lack of feature compatibility in public datasets. This limits generalization when facing distribution shifts, as well as transferability to different datasets. This study evaluates the suitability of different data preprocessing approaches for the detection of Portable Executable (PE) files with ML models. The preprocessing pipeline unifies EMBERv2 (2,381-dim) features datasets, trains paired models under two training setups: EMBER + BODMAS and EMBER + BODMAS + ERMDS. Regarding model evaluation, both EMBER + BODMAS and EMBER + BODMAS + ERMDS models are tested against TRITIUM, INFERNO and SOREL-20M. ERMDS is also used for testing for the EMBER + BODMAS setup.

Multimodal fusion requires integrating information from multiple sources that may conflict depending on context. Existing fusion approaches typically rely on static assumptions about source reliability, limiting their ability to resolve conflicts when a modality becomes unreliable due to situational factors such as sensor degradation or class-specific corruption. We introduce C$^2$MF, a context-specfic credibility-aware multimodal fusion framework that models per-instance source reliability using a Conditional Probabilistic Circuit (CPC). We formalize instance-level reliability through Context-Specific Information Credibility (CSIC), a KL-divergence-based measure computed exactly from the CPC. CSIC generalizes conventional static credibility estimates as a special case, enabling principled and adaptive reliability assessment. To evaluate robustness under cross-modal conflicts, we propose the Conflict benchmark, in which class-specific corruptions deliberately induce discrepancies between different modalities. Experimental results show that C$^2$MF improves predictive accuracy by up to 29% over static-reliability baselines in high-noise settings, while preserving the interpretability advantages of probabilistic circuit-based fusion.

Conditional density estimation (CDE) - recovering the full conditional distribution of a response given tabular covariates - is essential in settings with heteroscedasticity, multimodality, or asymmetric uncertainty. Recent tabular foundation models, such as TabPFN and TabICL, naturally produce predictive distributions, but their effectiveness as general-purpose CDE methods has not been systematically evaluated, unlike their performance for point prediction, which is well studied. We benchmark three tabular foundation model variants against a diverse set of parametric, tree-based, and neural CDE baselines on 39 real-world datasets, across training sizes from 50 to 20,000, using six metrics covering density accuracy, calibration, and computation time. Across all sample sizes, foundation models achieve the best CDE loss, log-likelihood, and CRPS on the large majority of datasets tested. Calibration is competitive at small sample sizes but, for some metrics and datasets, lags behind task-specific neural baselines at larger sample sizes, suggesting that post-hoc recalibration may be a valuable complement. In a photometric redshift case study using SDSS DR18, TabPFN exposed to 50,000 training galaxies outperforms all baselines trained on the full 500,000-galaxy dataset. Taken together, these results establish tabular foundation models as strong off-the-shelf conditional density estimators.

Mobile devices continuously interact with cellular base stations, generating massive volumes of signaling records that provide broad coverage for understanding human mobility. However, such records offer only coarse location cues (e.g., serving-cell identifiers) and therefore limit their direct use in applications that require high-precision GPS trajectories. This paper studies the Sig2GPS problem: reconstructing GPS trajectories from cellular signaling. Inspired by domain experts often lay the signaling trace on the map and sketch the corresponding GPS route, unlike conventional solutions that rely on complex multi-stage engineering pipelines or regress coordinates, Sig2GPS is reframed as an image-to-video generation task that directly operates in the map-visual domain: signaling traces are rendered on a map, and a video generation model is trained to draw a continuous GPS path. To support this paradigm, a paired signaling-to-trajectory video dataset is constructed to fine-tune an open-source video model, and a trajectory-aware reinforcement learning-based optimization method is introduced to improve generation fidelity via rewards. Experiments on large-scale real-world datasets show substantial improvements over strong engineered and learning-based baselines, while additional results on next GPS prediction indicate scalability and cross-city transferability. Overall, these results suggest that map-visual video generation provides a practical interface for trajectory data mining by enabling direct generation and refinement of continuous paths under map constraints.

Quantum machine learning offers the ability to capture complex correlations in high-dimensional feature spaces, crucial for the challenge of detecting beyond the Standard Model physics in collider events, along with the potential for unprecedented computational efficiency in future quantum processors. Near-term utilization of these benefits can be achieved by developing quantum-inspired algorithms for deployment in classical hardware to enable applications at the "edge" of current scientific experiments. This work demonstrates the use of tensor networks for real-time anomaly detection in collider detectors. A spaced matrix product operator (SMPO) is developed that provides sensitivity to a variety beyond the Standard Model benchmarks, and can be implemented in field programmable gate array hardware with resources and latency consistent with trigger deployment. The cascaded SMPO architecture is introduced as an SMPO variation that affords greater flexibility and efficiency in ways that are key to edge applications in resource-constrained environments. These results reveal the benefit and near-term feasibility of deploying quantum-inspired ML in high energy colliders.

The migration of Large Language Models (LLMs) from cloud clusters to edge devices promises enhanced privacy and offline accessibility, but this transition encounters a harsh reality: the physical constraints of mobile batteries, thermal limits, and, most importantly, memory constraints. To navigate this landscape, we constructed a reproducible experimental pipeline to profile the complex interplay between energy consumption, latency, and quality. Unlike theoretical studies, we captured granular power metrics across eight models ranging from 0.5B to 9B parameters without requiring root access, ensuring our findings reflect realistic user conditions. We harness this pipeline to conduct an empirical case study on a flagship Android device, the Samsung Galaxy S25 Ultra, establishing foundational hypotheses regarding the trade-offs between generation quality, performance, and resource consumption. Our investigation uncovered a counter-intuitive quantization-energy paradox. While modern importance-aware quantization successfully reduces memory footprints to fit larger models into RAM, we found it yields negligible energy savings compared to standard mixed-precision methods. This proves that for battery life, the architecture of the model, not its quantization scheme, is the decisive factor. We further identified that Mixture-of-Experts (MoE) architectures defy the standard size-energy trend, offering the storage capacity of a 7B model while maintaining the lower energy profile of a 1B to 2B model. Finally, an analysis of these multi-objective trade-offs reveals a pragmatic sweet spot of mid-sized models, such as Qwen2.5-3B, that effectively balance response quality with sustainable energy consumption.

The rapid growth of solar energy is reshaping power system operations and increasing the complexity of grid management. As photovoltaic (PV) capacity expands, short-term fluctuations in PV generation introduce substantial operational uncertainty. At the same time, solar power ramp events intensify risks of grid instability and unplanned outages due to sudden large power fluctuations. Accurate identification, forecasting and mitigation of solar ramp events are therefore critical to maintaining grid stability. In this study, we analyze two years of PV power production from 6434 PV stations at 15-minute resolution. We develop quantitative metrics to define solar ramp events and systematically characterize their occurrence, frequency, and magnitude at a national scale. Furthermore, we examine the meteorological drivers of ramp events, highlighting the role of mesoscale cloud systems. In particular, we observe that ramp-up events are typically associated with cloud dissipation during the morning, while ramp-down events commonly occur when cloud cover increases in the afternoon. Additionally, we adopt a recently developed spatiotemporal forecasting framework to evaluate both deterministic and probabilistic PV power forecasts derived from deep learning and physics-based models, including SolarSTEPS, SHADECast, IrradianceNet, and IFS-ENS. The results show that SHADECast is the most reliable model, achieving a CRPS 10.8% lower than that of SolarSTEPS at a two-hour lead time. Nonetheless, state-of-the-art nowcasting models struggle to capture ramp dynamics, with forecast RMSE increasing by up to 50% compared to normal operating conditions. Overall, these results emphasize the need for improved high-resolution spatiotemporal modelling to enhance ramp prediction skill and support the reliable integration of large-scale solar generation into power systems.

PQuantML is a new open-source, hardware-aware neural network model compression library tailored to end-to-end workflows. Motivated by the need to deploy performant models to environments with strict latency constraints, PQuantML simplifies training of compressed models by providing a unified interface to apply pruning and quantization, either jointly or individually. The library implements multiple pruning methods with different granularities, as well as fixed-point quantization with support for High-Granularity Quantization. We evaluate PQuantML on representative tasks such as the jet substructure classification, so-called jet tagging, an on-edge problem related to real-time LHC data processing. Using various pruning methods with fixed-point quantization, PQuantML achieves substantial parameter and bit-width reductions while maintaining accuracy. The resulting compression is further compared against existing tools, such as QKeras and HGQ.

Reliable machine-learning models in biomedical settings depend on accurate labels, yet annotating biomedical time-series data remains challenging. Algorithmic sample selection may support annotation, but evidence from studies involving real human annotators is scarce. Consequently, we compare three sample selection methods for annotation: random sampling (RND), farthest-first traversal (FAFT), and a graphical user interface-based method enabling exploration of complementary 2D visualizations (2DVs) of high-dimensional data. We evaluated the methods across four classification tasks in infant motility assessment (IMA) and speech emotion recognition (SER). Twelve annotators, categorized as experts or non-experts, performed data annotation under a limited annotation budget, and post-annotation experiments were conducted to evaluate the sampling methods. Across all classification tasks, 2DV performed best when aggregating labels across annotators. In IMA, 2DV most effectively captured rare classes, but also exhibited greater annotator-to-annotator label distribution variability resulting from the limited annotation budget, decreasing classification performance when models were trained on individual annotators' labels; in these cases, FAFT excelled. For SER, 2DV outperformed the other methods among expert annotators and matched their performance for non-experts in the individual-annotator setting. A failure risk analysis revealed that RND was the safest choice when annotator count or annotator expertise was uncertain, whereas 2DV had the highest risk due to its greater label distribution variability. Furthermore, post-experiment interviews indicated that 2DV made the annotation task more interesting and enjoyable. Overall, 2DV-based sampling appears promising for biomedical time-series data annotation, particularly when the annotation budget is not highly constrained.

We present ToothCraft, a diffusion-based model for the contextual generation of tooth crowns, trained on artificially created incomplete teeth. Building upon recent advancements in conditioned diffusion models for 3D shapes, we developed a model capable of an automated tooth crown completion conditioned on local anatomical context. To address the lack of training data for this task, we designed an augmentation pipeline that generates incomplete tooth geometries from a publicly available dataset of complete dental arches (3DS, ODD). By synthesising a diverse set of training examples, our approach enables robust learning across a wide spectrum of tooth defects. Experimental results demonstrate the strong capability of our model to reconstruct complete tooth crowns, achieving an intersection over union (IoU) of 81.8% and a Chamfer Distance (CD) of 0.00034 on synthetically damaged testing restorations. Our experiments demonstrate that the model can be applied directly to real-world cases, effectively filling in incomplete teeth, while generated crowns show minimal intersection with the opposing dentition, thus reducing the risk of occlusal interference. Access to the code, model weights, and dataset information will be available at: https://github.com/ikarus1211/VISAPP_ToothCraft

This paper investigates the relationship between utterance sentiment and language choice in English-Tamil code-switched text, using methods from machine learning and statistical modelling. We apply a fine-tuned XLM-RoBERTa model for token-level language identification on 35,650 romanized YouTube comments from the DravidianCodeMix dataset, producing per-utterance measurements of English proportion and language switch frequency. Linear regression analysis reveals that positive utterances exhibit significantly greater English proportion (34.3%) than negative utterances (24.8%), and mixed-sentiment utterances show the highest language switch frequency when controlling for utterance length. These findings support the hypothesis that emotional content demonstrably influences language choice in multilingual code-switching settings, due to socio-linguistic associations of prestige and identity with embedded and matrix languages.

Climbing is a multifaceted sport that combines physical demands and emotional and cognitive challenges. Ascent styles differ in fall distance with lead climbing involving larger falls than top rope climbing, which may result in different perceived risk and fear. In this study, we investigated the psychophysiological relationship between perceived fear and muscle activity in climbers using a combination of statistical modeling and deep learning techniques. We conducted an experiment with 19 climbers, collecting electromyography (EMG), electrocardiography (ECG) and arm motion data during lead and top rope climbing. Perceived fear ratings were collected for the different phases of the climb. Using a linear mixed-effects model, we analyzed the relationships between perceived fear and physiological measures. To capture the non-linear dynamics of this relationship, we extended our analysis to deep learning models and integrated random effects for a personalized modeling approach. Our results showed that random effects improved model performance of the mean squared error (MSE), mean absolute error (MAE) and root mean squared error (RMSE). The results showed that muscle fatigue correlates significantly with increased fear during \textit{lead climbing}. This study highlights the potential of combining statistical and deep learning approaches for modeling the interplay between psychological and physiological states during climbing.

Existing generative video compression methods use generative models only as post-hoc reconstruction modules atop conventional codecs. We propose \emph{Generative Video Codec} (GVC), a zero-shot framework that turns a pretrained video generative model into the codec itself: the transmitted bitstream directly specifies the generative decoding trajectory, with no retraining required. To enable this, we convert the deterministic rectified-flow ODE of modern video foundation models into an equivalent SDE at inference time, unlocking per-step stochastic injection points for codebook-driven compression. Building on this unified backbone, we instantiate three complementary conditioning strategies -- \emph{Image-to-Video} (I2V) with adaptive tail-frame atom allocation, \emph{Text-to-Video} (T2V) operating at near-zero side information as a pure generative prior, and \emph{First-Last-Frame-to-Video} (FLF2V) with boundary-sharing GOP chaining for dual-anchor temporal control. Together, these variants span a principled trade-off space between spatial fidelity, temporal coherence, and compression efficiency. Experiments on standard benchmarks show that GVC achieves high-quality reconstruction below 0.002\,bpp while supporting flexible bitrate control through a single hyperparameter.

Forgetting a subset in machine unlearning is rarely an isolated task. Often, retained samples that are closely related to the forget set can be unintentionally affected, particularly when they share correlated features from pretraining or exhibit strong semantic similarities. To address this challenge, we propose a novel two-phase optimization framework specifically designed to handle such retai-forget entanglements. In the first phase, an augmented Lagrangian method increases the loss on the forget set while preserving accuracy on less-related retained samples. The second phase applies a gradient projection step, regularized by the Wasserstein-2 distance, to mitigate performance degradation on semantically related retained samples without compromising the unlearning objective. We validate our approach through comprehensive experiments on multiple unlearning tasks, standard benchmark datasets, and diverse neural architectures, demonstrating that it achieves effective and reliable unlearning while outperforming existing baselines in both accuracy retention and removal fidelity.

Large Language Models (LLMs) have shown impressive capabilities across software engineering tasks, including question answering (QA). However, most studies and benchmarks focus on isolated functions or single-file snippets, overlooking the challenges of real-world program comprehension, which often spans multiple files and system-level dependencies. In this work, we introduce StackRepoQA, the first multi-project, repository-level question answering dataset constructed from 1,318 real developer questions and accepted answers across 134 open-source Java projects. Using this dataset, we systematically evaluate two widely used LLMs (Claude 3.5 Sonnet and GPT-4o) under both direct prompting and agentic configurations. We compare baseline performance with retrieval-augmented generation methods that leverage file-level retrieval and graph-based representations of structural dependencies. Our results show that LLMs achieve moderate accuracy at baseline, with performance improving when structural signals are incorporated. Nonetheless, overall accuracy remains limited for repository-scale comprehension. The analysis reveals that high scores often result from verbatim reproduction of Stack Overflow answers rather than genuine reasoning. To our knowledge, this is the first empirical study to provide such evidence in repository-level QA. We release StackRepoQA to encourage further research into benchmarks, evaluation protocols, and augmentation strategies that disentangle memorization from reasoning, advancing LLMs as reliable tool for repository-scale program comprehension.

Large Language Models (LLMs) deliver strong performance but incur high inference cost in real-world services, especially under workloads with repeated or near-duplicate queries across users and sessions. In this work, we propose MemBoost, a memory-boosted LLM serving framework that enables a lightweight model to reuse previously generated answers and retrieve relevant supporting information for cheap inference, while selectively escalating difficult or uncertain queries to a stronger model. Unlike standard retrieval-augmented generation, which primarily grounds a single response, MemBoost is designed for interactive settings by supporting answer reuse, continual memory growth, and cost-aware routing. Experiments across multiple models under simulated workloads show that MemBoost substantially reduces expensive large-model invocations and overall inference cost, while maintaining high answer quality comparable to the strong model baseline.

Converting a pretrained Transformer into a more efficient hybrid model through distillation offers a promising approach to reducing inference costs. However, achieving high-quality generation in distilled models requires careful joint design of both the student architecture and the distillation process. Many prior distillation works evaluate downstream multiple-choice benchmarks by ranking candidate answers with log-likelihood rather than requiring autoregressive generation, which can obscure important differences in model quality. For example, we show that a 7B parameter distilled model that nearly matches its teacher to within 0.2\,pp under log-likelihood scoring actually falls behind by 20.8\,pp when the model must generate answers autoregressively. We propose a Hybrid Kimi Delta Attention (Hybrid-KDA) architecture paired with GenDistill, a multi-stage distillation pipeline, and use generation-based evaluation throughout to guide design decisions. Applying this approach to Qwen3-0.6B, we systematically ablate six design axes: training objective, loss masking, training duration, dataset selection, parameter freezing, and architecture choice. We find that log-likelihood-based evaluation consistently underestimates the gap between teacher and student, and can in some cases reverse the ranking of design choices, meaning that conclusions drawn from perplexity-only evaluation may be misleading. Among the factors we study, dataset selection, completion-only masking, and freezing attention layers during post-training have the largest impact on generation quality. Our best Hybrid-KDA model retains 86--90\% of teacher accuracy on knowledge benchmarks while reducing KV cache memory by up to 75\% and improving time-to-first-token by 2--4$\times$ at 128K-token contexts.

Spectral optimizers such as Muon have recently shown strong empirical performance in large-scale language model training, but the source and extent of their advantage remain poorly understood. We study this question through the linear associative memory problem, a tractable model for factual recall in transformer-based models. In particular, we go beyond orthogonal embeddings and consider Gaussian inputs and outputs, which allows the number of stored associations to greatly exceed the embedding dimension. Our main result sharply characterizes the recovery rates of one step of Muon and SGD on the logistic regression loss under a power law frequency distribution. We show that the storage capacity of Muon significantly exceeds that of SGD, and moreover Muon saturates at a larger critical batch size. We further analyze the multi-step dynamics under a thresholded gradient approximation and show that Muon achieves a substantially faster initial recovery rate than SGD, while both methods eventually converge to the information-theoretic limit at comparable speeds. Experiments on synthetic tasks validate the predicted scaling laws. Our analysis provides a quantitative understanding of the signal amplification of Muon and lays the groundwork for establishing scaling laws across more practical language modeling tasks and optimizers.

Vision backbone networks play a central role in modern computer vision. Enhancing their efficiency directly benefits a wide range of downstream applications. To measure efficiency, many publications rely on MACs (Multiply Accumulate operations) as a predictor of execution time. In this paper, we experimentally demonstrate the shortcomings of such a metric, especially in the context of edge devices. By contrasting the MAC count and execution time of common architectural design elements, we identify key factors for efficient execution and provide insights to optimize backbone design. Based on these insights, we present LowFormer, a novel vision backbone family. LowFormer features a streamlined macro and micro design that includes Lowtention, a lightweight alternative to Multi-Head Self-Attention. Lowtention not only proves more efficient, but also enables superior results on ImageNet. Additionally, we present an edge GPU version of LowFormer, that can further improve upon its baseline's speed on edge GPU and desktop GPU. We demonstrate LowFormer's wide applicability by evaluating it on smaller image classification datasets, as well as adapting it to several downstream tasks, such as object detection, semantic segmentation, image retrieval, and visual object tracking. LowFormer models consistently achieve remarkable speed-ups across various hardware platforms compared to recent state-of-the-art backbones. Code and models are available at https://github.com/altair199797/LowFormer/blob/main/Beyond_MACs.md.

We adapt the analysis of policy gradient for continuous time $k$-armed stochastic bandits by Lattimore (2026) to the standard discrete time setup. As in continuous time, we prove that with learning rate $η= O(Δ_{\min}^2/(Δ_{\max} \log(n)))$ the regret is $O(k \log(k) \log(n) / η)$ where $n$ is the horizon and $Δ_{\min}$ and $Δ_{\max}$ are the minimum and maximum gaps.

Background: The identification of optimal signal detection methods is hindered by the lack of reliable reference datasets. Existing datasets do not capture when adverse events (AEs) are officially recognized by regulatory authorities, preventing restriction of analyses to pre-confirmation periods and limiting evaluation of early detection performance. This study addresses this gap by developing a time-indexed reference dataset for the European Union (EU), incorporating the timing of AE inclusion in product labels along with regulatory metadata. Methods: Current and historical Summaries of Product Characteristics (SmPCs) for all centrally authorized products (n=1,513) were retrieved from the EU Union Register of Medicinal Products (data lock: 15 December 2025). Section 4.8 was extracted and processed using DeepSeek V3 to identify AEs. Regulatory metadata, including labelling changes, were programmatically extracted. Time indexing was based on the date of AE inclusion in the SmPC. Results: The database includes 17,763 SmPC versions spanning 1995-2025, comprising 125,026 drug-AE associations. The time-indexed reference dataset, restricted to active products, included 1,479 medicinal products and 110,823 drug-AE associations. Most AEs were identified pre-marketing (74.5%) versus post-marketing (25.5%). Safety updates peaked around 2012. Gastrointestinal, skin, and nervous system disorders were the most represented System Organ Classes. Drugs had a median of 48 AEs across 14 SOCs. Conclusions: The proposed dataset addresses a critical gap in pharmacovigilance by incorporating temporal information on AE recognition for the EU, supporting more accurate assessment of signal detection performance and facilitating methodological comparisons across analytical approaches.

Buffer zones are essential in production systems to decouple sequential processes. In dense floor storage environments, such as space-constrained brownfield facilities, manual operation is increasingly challenged by severe labor shortages and rising operational costs. Automating these zones requires solving the Buffer Storage, Retrieval, and Reshuffling Problem (BSRRP). While previous work has addressed scenarios where the focus is limited to reshuffling and retrieving a fixed set of items, real-world manufacturing necessitates an adaptive approach that also incorporates arriving unit loads. This paper introduces the Multi-AMR BSRRP, coordinating a robot fleet to manage concurrent reshuffling, alongside time-windowed storage and retrieval tasks, within a shared floor area. We formulate a Binary Integer Programming (IP) model to obtain exact solutions for benchmarking purposes. As the problem is NP-hard, rendering exact methods computationally intractable for industrial scales, we propose a hierarchical heuristic. This approach decomposes the problem into an A* search for task-level sequence planning of unit load placements, and a Constraint Programming (CP) approach for multi-robot coordination and scheduling. Experiments demonstrate orders-of-magnitude computation time reductions compared to the exact formulation. These results confirm the heuristic's viability as responsive control logic for high-density production environments.

How does the extent to which a model is open or closed impact the scientific inferences that can be drawn from research that involves it? In this paper, we analyze how restrictions on information about model construction and deployment threaten reliable inference. We argue that current closed models are generally ill-suited for scientific purposes, with some notable exceptions, and discuss ways in which the issues they present to reliable inference can be resolved or mitigated. We recommend that when models are used in research, potential threats to inference should be systematically identified along with the steps taken to mitigate them, and that specific justifications for model selection should be provided.

We propose Process-Aware Policy Optimization (PAPO), a method that integrates process-level evaluation into Group Relative Policy Optimization (GRPO) through decoupled advantage normalization, to address two limitations of existing reward designs. Outcome reward models (ORM) evaluate only final-answer correctness, treating all correct responses identically regardless of reasoning quality, and gradually lose the advantage signal as groups become uniformly correct. Process reward models (PRM) offer richer supervision, but directly using PRM scores causes reward hacking, where models exploit verbosity to inflate scores while accuracy collapses. PAPO resolves both by composing the advantage from an outcome component Aout, derived from ORM and normalized over all responses, and a process component Aproc, derived from a rubric-based PRM and normalized exclusively among correct responses. This decoupled design ensures that Aout anchors training on correctness while Aproc differentiates reasoning quality without distorting the outcome signal. Experiments across multiple model scales and six benchmarks demonstrate that PAPO consistently outperforms ORM, reaching 51.3% vs.\ 46.3% on OlympiadBench while continuing to improve as ORM plateaus and declines.

Since the earliest stages of human civilization, advances in technology have been tightly linked to our ability to understand and predict the mechanical behavior of materials. In recent years, this challenge has increasingly been framed within the broader paradigm of data-driven scientific discovery, where governing laws are inferred directly from observations. However, existing methods require either stress-strain pairs or full-field displacement measurements, which are often inaccessible in practice. We introduce Neural-DFEM, a method that enables unsupervised discovery of hyperelastic material laws even from partial observations, such as boundary-only measurements. The method embeds a differentiable finite element solver within the learning loop, directly linking candidate energy functionals to available measurements. To guarantee thermodynamic consistency and mathematical well-posedness throughout training, the method employs Hyperelastic Neural Networks, a novel structure-preserving neural architecture that enforces frame indifference, material symmetry, polyconvexity, and coercivity by design. The resulting framework enables robust material model discovery in both two- and three-dimensional settings, including scenarios with boundary-only measurements. Neural-DFEM allows for generalization across geometries and loading conditions, and exhibits unprecedented accuracy and strong resilience to measurement noise. Our results demonstrate that reliable identification of material laws is achievable even under partial observability when strong physical inductive biases are embedded in the learning architecture.

As Large Language Models (LLMs) expand across multilingual domains, evaluating their performance in under-represented languages becomes increasingly important. European Portuguese (pt-PT) is particularly affected, as existing training data and benchmarks are mainly in Brazilian Portuguese (pt-BR). To address this, we introduce ALBA, a linguistically grounded benchmark designed from the ground up to assess LLM proficiency in linguistic-related tasks in pt-PT across eight linguistic dimensions, including Language Variety, Culture-bound Semantics, Discourse Analysis, Word Plays, Syntax, Morphology, Lexicology, and Phonetics and Phonology. ALBA is manually constructed by language experts and paired with an LLM-as-a-judge framework for scalable evaluation of pt-PT generated language. Experiments on a diverse set of models reveal performance variability across linguistic dimensions, highlighting the need for comprehensive, variety-sensitive benchmarks that support further development of tools in pt-PT.

Despite recent advances, efficient and robust turn-taking detection remains a significant challenge in industrial-grade Voice AI agent deployments. Many existing systems rely solely on acoustic or semantic cues, leading to suboptimal accuracy and stability, while recent attempts to endow large language models with full-duplex capabilities require costly full-duplex data and incur substantial training and deployment overheads, limiting real-time performance. In this paper, we propose JAL-Turn, a lightweight and efficient speech-only turn-taking framework that adopts a joint acoustic-linguistic modeling paradigm, in which a cross-attention module adaptively integrates pre-trained acoustic representations with linguistic features to support low-latency prediction of hold vs shift states. By sharing a frozen ASR encoder, JAL-Turn enables turn-taking prediction to run fully in parallel with speech recognition, introducing no additional end-to-end latency or computational overhead. In addition, we introduce a scalable data construction pipeline that automatically derives reliable turn-taking labels from large-scale real-world dialogue corpora. Extensive experiments on public multilingual benchmarks and an in-house Japanese customer-service dataset show that JAL-Turn consistently outperforms strong state-of-the-art baselines in detection accuracy while maintaining superior real-time performance.

Existing methods for text-to-CAD generation either operate in a single pass with no geometric verification or rely on lossy visual feedback that cannot resolve dimensional errors. We present CADSmith, a multi-agent pipeline that generates CadQuery code from natural language. It then undergoes an iterative refinement process through two nested correction loops: an inner loop that resolves execution errors and an outer loop grounded in programmatic geometric validation. The outer loop combines exact measurements from the OpenCASCADE kernel (bounding box dimensions, volume, solid validity) with holistic visual assessment from an independent vision-language model Judge. This provides both the numerical precision and the high-level shape awareness needed to converge on the correct geometry. The system uses retrieval-augmented generation over API documentation rather than fine-tuning, maintaining a current database as the underlying CAD library evolves. We evaluate on a custom benchmark of 100 prompts in three difficulty tiers (T1 through T3) with three ablation configurations. Against a zero-shot baseline, CADSmith achieves a 100% execution rate (up from 95%), improves the median F1 score from 0.9707 to 0.9846, the median IoU from 0.8085 to 0.9629, and reduces the mean Chamfer Distance from 28.37 to 0.74, demonstrating that closed-loop refinement with programmatic geometric feedback substantially improves the quality and reliability of LLM-generated CAD models.

Despite rapid progress in open large language models (LLMs), European Portuguese (pt-PT) remains underrepresented in both training data and native evaluation, with machine-translated benchmarks likely missing the variant's linguistic and cultural nuances. We introduce AMALIA, a fully open LLM that prioritizes pt-PT by using more high-quality pt-PT data during both the mid- and post-training stages. To evaluate pt-PT more faithfully, we release a suite of pt-PT benchmarks that includes translated standard tasks and four new datasets targeting pt-PT generation, linguistic competence, and pt-PT/pt-BR bias. Experiments show that AMALIA matches strong baselines on translated benchmarks while substantially improving performance on pt-PT-specific evaluations, supporting the case for targeted training and native benchmarking for European Portuguese.

Clinical notes contain valuable unstructured information. Named entity recognition (NER) enables the automatic extraction of medical concepts; however, benchmarks for Portuguese remain scarce. In this study, we aimed to evaluate BERT-based models and large language models (LLMs) for clinical NER in Portuguese and to test strategies for addressing multilabel imbalance. We compared BioBERTpt, BERTimbau, ModernBERT, and mmBERT with LLMs such as GPT-5 and Gemini-2.5, using the public SemClinBr corpus and a private breast cancer dataset. Models were trained under identical conditions and evaluated using precision, recall, and F1-score. Iterative stratification, weighted loss, and oversampling were explored to mitigate class imbalance. The mmBERT-base model achieved the best performance (micro F1 = 0.76), outperforming all other models. Iterative stratification improved class balance and overall performance. Multilingual BERT models, particularly mmBERT, perform strongly for Portuguese clinical NER and can run locally with limited computational resources. Balanced data-splitting strategies further enhance performance.

Connectivity structure shapes neural computation, but inferring this structure from population recordings is degenerate: multiple connectivity structures can generate identical dynamics. Recent work uses low-rank recurrent neural networks (lrRNNs) to infer low-dimensional latent dynamics and connectivity structure from observed activity, enabling a mechanistic interpretation of the dynamics. However, standard approaches for training lrRNNs can recover spurious structures irrelevant to the underlying dynamics. We first characterize the identifiability of connectivity structures in lrRNNs and determine conditions under which a unique solution exists. Then, to find such solutions, we develop an inference framework based on maximum entropy and continuous normalizing flows (CNFs), trained via flow matching. Instead of estimating a single connectivity matrix, our method learns the maximally unbiased distribution over connection weights consistent with observed dynamics. This approach captures complex yet necessary distributions such as heavy-tailed connectivity found in empirical data. We validate our method on synthetic datasets with connectivity structures that generate multistable attractors, limit cycles, and ring attractors, and demonstrate its applicability in recordings from rat frontal cortex during decision-making. Our framework shifts circuit inference from recovering connectivity to identifying which connectivity structures are computationally required, and which are artifacts of underconstrained inference.

Existing research has identified three structural performance bottlenecks in AI research agents: (1) synchronous single-GPU execution constrains sample throughput, limiting the benefit of search; (2) a generalization gap where validation-based selection causes performance to degrade over extended search horizons; and (3) the limited capability of fixed, single-turn LLM operators imposes a ceiling on search performance. We introduce AIRA$_2$, which addresses these bottlenecks through three architectural choices: an asynchronous multi-GPU worker pool that increases experiment throughput linearly; a Hidden Consistent Evaluation protocol that delivers a reliable evaluation signal; and ReAct agents that dynamically scope their actions and debug interactively. On MLE-bench-30, AIRA$_2$ achieves a mean Percentile Rank of 71.8% at 24 hours - surpassing the previous best of 69.9% - and steadily improves to 76.0% at 72 hours. Ablation studies reveal that each component is necessary and that the "overfitting" reported in prior work was driven by evaluation noise rather than true data memorization.

Multimodal Large Language Models (MLLMs) power platforms like ChatGPT, Gemini, and Copilot, enabling richer interactions with text, images, and videos. These heterogeneous workloads introduce additional inference stages, such as vision preprocessing and encoding, that inflate latency and memory demand. Existing LLM serving systems, optimized for text-only workloads, fail under multimodality: large requests (e.g., videos) monopolize resources, causing severe head-of-line blocking and performance degradation. Our key insight is that multimodal requests differ by orders of magnitude in resource demands, which we capture through a simple abstraction: videos behave like rocks, images like pebbles, and text like sand. We design RPS-Serve, a modality-aware scheduler that lets sand flow quickly through pebbles and rocks, ensuring interactive responsiveness while avoiding starvation. RPS-Serve classifies requests, prioritizes them dynamically, and applies aging to avoid starvation. Evaluation across state-of-the-art MLLMs shows that RPS-Serve reduces, on average, time-to-first-token (TTFT) by 54% overall, and by 78.5% for latency-critical requests, compared to current systems. RPS-Serve delivers LLM-like responsiveness for MLLMs, with modality-aware scheduling and by making the most efficient use of the available resources.

Reentrancy remains one of the most critical classes of vulnerabilities in Ethereum smart contracts, yet widely used detection tools and datasets continue to reflect outdated patterns and obsolete Solidity versions. This paper adopts a dependability-oriented perspective on reentrancy detection in Solidity 0.8+, assessing how reliably state-of-the-art static analyzers and AI-based techniques operate on modern code by putting them to the test on two fronts. We construct two manually verified benchmarks: an Aggregated Benchmark of 432 real-world contracts, consolidated and relabeled from prior datasets, and a Reentrancy Scenarios Dataset (RSD) of \chadded{143} handcrafted minimal working examples designed to isolate and stress-test individual reentrancy patterns. We then evaluate 12 formal-methods-based tools, 10 machine-learning models, and 9 large language models. On the Aggregated Benchmark, traditional tools and ML models achieve up to 0.87 F1, while the best LLMs reach 0.96 in a zero-shot setting. On the RSD, most tools fail on multiple scenarios, the top performer achieving an F1 of 0.76, whereas the strongest model attains 0.82. Overall, our results indicate that leading LLMs outperform the majority of existing detectors, highlighting concerning gaps in the robustness and maintainability of current reentrancy-analysis tools.

Quantum language models have shown competitive performance on sequential tasks, yet whether trained quantum circuits exploit genuinely quantum resources -- or merely embed classical computation in quantum hardware -- remains unknown. Prior work has evaluated these models through endpoint metrics alone, without examining the memory strategies they actually learn internally. We introduce the first mechanistic interpretability study of quantum language models, combining causal gate ablation, entanglement tracking, and density-matrix interchange interventions on a controlled long-range dependency task. We find that single-qubit models are exactly classically simulable and converge to the same geometric strategy as matched classical baselines, while two-qubit models with entangling gates learn a representationally distinct strategy that encodes context in inter-qubit entanglement -- confirmed by three independent causal tests (p < 0.0001, d = 0.89). On real quantum hardware, only the classical geometric strategy survives device noise; the entanglement strategy degrades to chance. These findings open mechanistic interpretability as a tool for the science of quantum language models and reveal a noise-expressivity tradeoff governing which learned strategies survive deployment.

For frontier experiments operating at the edge of detectability, instrument design directly determines the probability of discovery. We introduce Conditional Neural Bayes Ratio Estimation (cNBRE), which extends neural Bayes ratio estimation by conditioning on design parameters, enabling a single trained network to estimate Bayes factors across a continuous design space. Applied to 21-cm radio cosmology with simulations representative of the REACH experiment, the amortised nature of cNBRE enables systematic design space exploration that would be intractable with traditional point-wise methods, while recovering established physical relationships. The analysis demonstrates a ~20 percentage point variation in detection probability with antenna orientation for a single night of observation, a design decision that would be trivial to implement if determined prior to antenna construction. This framework enables efficient, globally-informed experimental design optimisation for a wide range of scientific applications.

Generative AI (GenAI) is playing an increasingly important role in open source software (OSS). Beyond completing code and documentation, GenAI is increasingly involved in issues, pull requests, code reviews, and security reports. Yet, cheaper generation does not mean cheaper review - and the resulting maintenance burden has pushed OSS projects to experiment with GenAI-specific rules in contribution guidelines, security policies, and repository instructions, even including a total ban on AI-assisted contributions. However, governing GenAI in OSS is far more than a ban-or-not question. The responses remain scattered, with neither a shared governance framework in practice nor a systematic understanding in research. Therefore, in this paper, we conduct a multi-stage analysis on various qualitative materials related to GenAI governance retrieved from 67 highly visible OSS projects. Our analysis identifies recurring concerns across contribution workflows, derives three governance orientations, and maps out 12 governance strategies and their implementation patterns. We show that governing GenAI in OSS extends well beyond banning - it requires coordinated responses across accountability, verification, review capacity, code provenance, and platform infrastructure. Overall, our work distills dispersed community practices into a structured overview, providing a conceptual baseline for researchers and a practical reference for maintainers and platform designers.

Privacy-preserving medical inference must balance data locality, diagnostic reliability, and deployment efficiency. This paper presents EcoFair, a simulated vertically partitioned inference framework for dermatological diagnosis in which raw image and tabular data remain local and only modality-specific embeddings are transmitted for server-side multimodal fusion. EcoFair introduces a lightweight-first routing mechanism that selectively activates a heavier image encoder when local uncertainty or metadata-derived clinical risk indicates that additional computation is warranted. The routing decision combines predictive uncertainty, a safe--danger probability gap, and a tabular neurosymbolic risk score derived from patient age and lesion localisation. Experiments on three dermatology benchmarks show that EcoFair can substantially reduce edge-side inference energy in representative model pairings while remaining competitive in classification performance. The results further indicate that selective routing can improve subgroup-sensitive malignant-case behaviour in representative settings without modifying the global training objective. These findings position EcoFair as a practical framework for privacy-preserving and energy-aware medical inference under edge deployment constraints.

Real time sensor based applications in pervasive computing require edge deployable models to ensure low latency privacy and efficient interaction. A prime example is sensor based human activity recognition where models must balance accuracy with stringent resource constraints. Yet many deep learning approaches treat temporal sensor signals as black box sequences overlooking spectral temporal structure while demanding excessive computation. We present SPECTRA a deployment first co designed spectral temporal architecture that integrates short time Fourier transform STFT feature extraction depthwise separable convolutions and channel wise self attention to capture spectral temporal dependencies under real edge runtime and memory constraints. A compact bidirectional GRU with attention pooling summarizes within window dynamics at low cost reducing downstream model burden while preserving accuracy. Across five public HAR datasets SPECTRA matches or approaches larger CNN LSTM and Transformer baselines while substantially reducing parameters latency and energy. Deployments on a Google Pixel 9 smartphone and an STM32L4 microcontroller further demonstrate end to end deployable realtime private and efficient HAR.

Explainability and fairness have mainly been considered separately, with recent exceptions trying the explain the sources of unfairness. This paper shows that the Shapley value can be used to both define and explain unfairness, under standard group fairness criteria. This offers an integrated framework to estimate and derive inference on unfairness as-well-as the features that contribute to it. Our framework can also be extended from Shapley values to the family of Efficient-Symmetric-Linear (ESL) values, some of which offer more robust definitions of fairness, and shorter computation times. An illustration is run on the Census Income dataset from the UCI Machine Learning Repository. Our approach shows that ``Age", ``Number of hours" and ``Marital status" generate gender unfairness, using shorter computation time than traditional Bootstrap tests.

Electrocardiograms (ECGs) are among the most widely available clinical signals and play a central role in cardiovascular diagnosis. While recent foundation models (FMs) have shown promise for learning transferable ECG representations, most existing pretraining approaches treat leads as independent channels and fail to explicitly leverage their strong structural redundancy. We introduce the latent attention masked autoencoder (LAMAE) FM that directly exploits this structure by learning cross-lead connection mechanisms during self-supervised pretraining. Our approach models higher-order interactions across leads through latent attention, enabling permutation-invariant aggregation and adaptive weighting of lead-specific representations. We provide empirical evidence on the Mimic-IV-ECG database that leveraging the cross-lead connection constitutes an effective form of structural supervision, improving representation quality and transferability. Our method shows strong performance in predicting ICD-10 codes, outperforming independent-lead masked modeling and alignment-based baselines.

Developing and evaluating distributed inference algorithms remains difficult due to the lack of standardized tools for modeling heterogeneous devices and networks. Existing studies often rely on ad-hoc testbeds or proprietary infrastructure, making results hard to reproduce and limiting exploration of hypothetical hardware or network configurations. We present UNIFERENCE, a discrete-event simulation (DES) framework designed for developing, benchmarking, and deploying distributed AI models within a unified environment. UNIFERENCE models device and network behavior through lightweight logical processes that synchronize only on communication primitives, eliminating rollbacks while preserving the causal order. It integrates seamlessly with PyTorch Distributed, enabling the same codebase to transition from simulation to real deployment. Our evaluation demonstrates that UNIFERENCE profiles runtime with up to 98.6% accuracy compared to real physical deployments across diverse backends and hardware setups. By bridging simulation and deployment, UNIFERENCE provides an accessible, reproducible platform for studying distributed inference algorithms and exploring future system designs, from high-performance clusters to edge-scale devices. The framework is open-sourced at https://github.com/Dogacel/Uniference.

DNA sequence classification requires not only high predictive accuracy but also the ability to uncover latent site interactions, combinatorial regulation, and epistasis-like higher-order dependencies. Although the standard Transformer provides strong global modeling capacity, its softmax attention is continuous, dense, and weakly constrained, making it better suited for information routing than explicit structure discovery. In this paper, we propose a Boltzmann-machine-enhanced Transformer for DNA sequence classification. Built on multi-head attention, the model introduces structured binary gating variables to represent latent query-key connections and constrains them with a Boltzmann-style energy function. Query-key similarity defines local bias terms, learnable pairwise interactions capture synergy and competition between edges, and latent hidden units model higher-order combinatorial dependencies. Since exact posterior inference over discrete gating graphs is intractable, we use mean-field variational inference to estimate edge activation probabilities and combine it with Gumbel-Softmax to progressively compress continuous probabilities into near-discrete gates while preserving end-to-end differentiability. During training, we jointly optimize classification and energy losses, encouraging the model to achieve accurate prediction while favoring low-energy, stable, and interpretable structures. We further derive the framework from the energy function and variational free energy to the mean-field fixed-point equations, Gumbel-Softmax relaxation, and the final joint objective. The proposed framework provides a unified view of integrating Boltzmann machines, differentiable discrete optimization, and Transformers for structured learning on biological sequences.

In this paper, we present a Koopman autoencoder-based least-squares policy iteration (KAE-LSPI) algorithm in reinforcement learning (RL). The KAE-LSPI algorithm is based on reformulating the so-called least-squares fixed-point approximation method in terms of extended dynamic mode decomposition (EDMD), thereby enabling automatic feature learning via the Koopman autoencoder (KAE) framework. The approach is motivated by the lack of a systematic choice of features or kernels in linear RL techniques. We compare the KAE-LSPI algorithm with two previous works, the classical least-squares policy iteration (LSPI) and the kernel-based least-squares policy iteration (KLSPI), using stochastic chain walk and inverted pendulum control problems as examples. Unlike previous works, no features or kernels need to be fixed a priori in our approach. Empirical results show the number of features learned by the KAE technique remains reasonable compared to those fixed in the classical LSPI algorithm. The convergence to an optimal or a near-optimal policy is also comparable to the other two methods.

Process anomaly detection is an important application of process mining for identifying deviations from the normal behavior of a process. Neural network-based methods have recently been applied to this task, learning directly from event logs without requiring a predefined process model. However, since anomaly detection is a purely statistical task, these models fail to incorporate human domain knowledge. As a result, rare but conformant traces are often misclassified as anomalies due to their low frequency, which limits the effectiveness of the detection process. Recent developments in the field of neuro-symbolic AI have introduced Logic Tensor Networks (LTN) as a means to integrate symbolic knowledge into neural networks using real-valued logic. In this work, we propose a neuro-symbolic approach that integrates domain knowledge into neural anomaly detection using LTN and Declare constraints. Using autoencoder models as a foundation, we encode Declare constraints as soft logical guiderails within the learning process to distinguish between anomalous and rare but conformant behavior. Evaluations on synthetic and real-world datasets demonstrate that our approach improves F1 scores even when as few as 10 conformant traces exist, and that the choice of Declare constraint and by extension human domain knowledge significantly influences performance gains.

Can an expensive AI model effectively direct a cheap one to solve software engineering tasks? We study this question by introducing ManagerWorker, a two-agent pipeline where an expensive "manager" model (text-only, no code execution) analyzes issues, dispatches exploration tasks, and reviews implementations, while a cheap "worker" model (with full repo access) executes code changes. We evaluate on 200 instances from SWE-bench Lite across five configurations that vary the manager-worker relationship, pipeline complexity, and model pairing. Our findings reveal both the promise and the limits of multi-agent direction: (1) a strong manager directing a weak worker (62%) matches a strong single agent (60%) at a fraction of the strong-model token usage, showing that expensive reasoning can substitute for expensive execution; (2) a weak manager directing a weak worker (42%) performs worse than the weak agent alone (44%), demonstrating that the directing relationship requires a genuine capability gap--structure without substance is pure overhead; (3) the manager's value lies in directing, not merely reviewing--a minimal review-only loop adds just 2pp over the baseline, while structured exploration and planning add 11pp, showing that active direction is what makes the capability gap productive; and (4) these behaviors trace to a single root cause: current models are trained as monolithic agents, and splitting them into director/worker roles fights their training distribution. The pipeline succeeds by designing around this mismatch--keeping each model close to its trained mode (text generation for the manager, tool use for the worker) and externalizing organizational structure to code. This diagnosis points to concrete training gaps: delegation, scoped execution, and mode switching are skills absent from current training data.

Fair data pre-processing is a widely used strategy for mitigating bias in machine learning. A promising line of research focuses on calibrating datasets to satisfy a designed fairness policy so that sensitive attributes influence outcomes only through clearly specified legitimate causal pathways. While effective on clean and information-rich data, these methods often break down in real-world scenarios with imperfect attribute spaces, where decision-relevant factors may be deemed unusable or even missing. To address this gap, we propose LatentPre, a novel framework that enables principled and robust fair data processing in practical settings. Instead of relying solely on observed attributes, LatentPre augments the fairness policy with latent attributes that capture essential but subtle signals, enabling the framework to operate as if the attribute space were perfect. These latent attributes are strategically introduced to guarantee identifiability and are estimated using a tailored expectation-maximization paradigm. The raw data is then carefully refined to conform to this latent-augmented policy, effectively removing biased patterns while preserving justifiable ones. Extensive experiments demonstrate that LatentPre consistently achieves strong fairness-utility trade-offs across diverse scenarios, advancing practical fairness-aware data management.

Automatically verifying climate-related claims against scientific literature is a challenging task, complicated by the specialised nature of scholarly evidence and the diversity of rhetorical strategies underlying climate disinformation. ClimateCheck 2026 is the second iteration of a shared task addressing this challenge, expanding on the 2025 edition with tripled training data and a new disinformation narrative classification task. Running from January to February 2026 on the CodaBench platform, the competition attracted 20 registered participants and 8 leaderboard submissions, with systems combining dense retrieval pipelines, cross-encoder ensembles, and large language models with structured hierarchical reasoning. In addition to standard evaluation metrics (Recall@K and Binary Preference), we adapt an automated framework to assess retrieval quality under incomplete annotations, exposing systematic biases in how conventional metrics rank systems. A cross-task analysis further reveals that not all climate disinformation is equally verifiable, potentially implicating how future fact-checking systems should be designed.

Human mesh recovery from single images remains challenging due to inherent depth ambiguity and limited generalization across domains. While recent methods combine regression and optimization approaches, they struggle with poor initialization for test-time refinement and inefficient parameter updates during optimization. We propose a novel meta-learning framework that trains models to produce optimization-friendly initializations while incorporating uncertainty-aware adaptive updates during test-time refinement. Our approach introduces three key innovations: (1) a meta-learning strategy that simulates test-time optimization during training to learn better parameter initializations, (2) a selective parameter caching mechanism that identifies and freezes converged joints to reduce computational overhead, and (3) distribution-based adaptive updates that sample parameter changes from learned distributions, enabling robust exploration while quantifying uncertainty. Additionally, we employ stochastic approximation techniques to handle intractable gradients in complex loss landscapes. Extensive experiments on standard benchmarks demonstrate that our method achieves state-of-the-art performance, reducing MPJPE by 10.3 on 3DPW and 8.0 on Human3.6M compared to strong baselines. Our approach shows superior domain adaptation capabilities with minimal performance degradation across different environmental conditions, while providing meaningful uncertainty estimates that correlate with actual prediction errors. Combining meta-learning and adaptive optimization enables accurate mesh recovery and robust generalization to challenging scenarios.

Long-term traffic modelling is fundamental to transport planning, but existing approaches often trade off interpretability, transferability, and predictive accuracy. Classical travel demand models provide behavioural structure but rely on strong assumptions and extensive calibration, whereas generic deep learning models capture complex patterns but often lack theoretical grounding and spatial transferability, limiting their usefulness for long-term planning applications. We propose DeepDemand, a theory-informed deep learning framework that embeds key components of travel demand theory to predict long-term highway traffic volumes using external socioeconomic features and road-network structure. The framework integrates a competitive two-source Dijkstra procedure for local origin-destination (OD) region extraction and OD pair screening with a differentiable architecture modelling OD interactions and travel-time deterrence. The model is evaluated using eight years (2017-2024) of observations on the UK strategic road network, covering 5088 highway segments. Under random cross-validation, DeepDemand achieves an R2 of 0.718 and an MAE of 7406 vehicles, outperforming linear, ridge, random forest, and gravity-style baselines. Performance remains strong under spatial cross-validation (R2 = 0.665), indicating good geographic transferability. Interpretability analysis reveals a stable nonlinear travel-time deterrence pattern, key socioeconomic drivers of demand, and polycentric OD interaction structures aligned with major employment centres and transport hubs. These results highlight the value of integrating transport theory with deep learning for interpretable highway traffic modelling and practical planning applications.

Clinicians often need to retrieve patient-specific information from electronic health records (EHRs), a task that is time-consuming and error-prone. We present a locally deployable Clinical Contextual Question Answering (CCQA) framework that answers clinical questions directly from EHRs without external data transfer. Open-source large language models (LLMs) ranging from 4B to 70B parameters were benchmarked under fully offline conditions using 1,664 expert-annotated question-answer pairs derived from records of 183 patients. The dataset consisted predominantly of Finnish clinical text. In free-text generation, Llama-3.1-70B achieved 95.3% accuracy and 97.3% consistency across semantically equivalent question variants, while the smaller Qwen3-30B-A3B-2507 model achieved comparable performance. In a multiple-choice setting, models showed similar accuracy but variable calibration. Low-precision quantization (4-bit and 8-bit) preserved predictive performance while reducing GPU memory requirements and improving deployment feasibility. Clinical evaluation identified clinically significant errors in 2.9% of outputs, and semantically equivalent questions occasionally yielded discordant responses, including instances where one formulation was correct and the other contained a clinically significant error (0.96% of cases). These findings demonstrate that locally hosted open-source LLMs can accurately retrieve patient-specific information from EHRs using natural-language queries, while highlighting the need for validation and human oversight in clinical deployment.

Efficiently characterizing quantum dot (QD) devices is a critical bottleneck when scaling quantum processors based on confined spins. Measuring high-resolution charge stability diagrams (or CSDs, data maps which crucially define the occupation of QDs) is time-consuming, particularly in emerging architectures where CSDs must be acquired with remote sensors that cannot probe the charge of the relevant dots directly. In this work, we present a generative approach to accelerate acquisition by reconstructing full CSDs from sparse measurements, using a conditional diffusion model. We evaluate our approach using two experimentally motivated masking strategies: uniform grid-based sampling, and line-cut sweeps. Our lightweight architecture, trained on approximately 9,000 examples, successfully reconstructs CSDs, maintaining key physically important features such as charge transition lines, from as little as 4\% of the total measured data. We compare the approach to interpolation methods, which fail when the task involves reconstructing large unmeasured regions. Our results demonstrate that generative models can significantly reduce the characterization overhead for quantum devices, and provides a robust path towards an experimental implementation.

Parliamentary debate constitutes a central arena of political power, shaping legislative outcomes and public discourse. Incivility within this arena signals political polarization and institutional conflict. This study presents a systematic investigation of incivility in the German Bundestag by examining calls to order (CtO; plural: CtOs) as formal indicators of norm violations. Despite their relevance, CtOs have received little systematic attention in parliamentary research. We introduce a rule-based method for detecting and annotating CtOs in parliamentary speeches and present a novel dataset of German parliamentary debates spanning 72 years that includes annotated CtO instances. Additionally, we develop the first classification system for CtO triggers and analyze the factors associated with their occurrence. Our findings show that, despite formal regulations, the issuance of CtOs is partly subjective and influenced by session presidents and parliamentary dynamics, with certain individuals disproportionately affected. An insult towards individuals is the most frequent cause of CtO. In general, male members and those belonging to opposition parties receive more calls to order than their female and coalition-party counterparts. Most CtO triggers were detected in speeches dedicated to governmental affairs and actions of the presidency. The CtO triggers dataset is available at: https://github.com/kalawinka/cto_analysis.

Recent research on vision backbone architectures has predominantly focused on optimizing efficiency for hardware platforms with high parallel processing capabilities. This category increasingly includes embedded systems such as mobile phones and embedded AI accelerator modules. In contrast, CPUs do not have the possibility to parallelize operations in the same manner, wherefore models benefit from a specific design philosophy that balances amount of operations (MACs) and hardware-efficient execution by having high MACs per second (MACpS). In pursuit of this, we investigate two modifications to standard convolutions, aimed at reducing computational cost: grouping convolutions and reducing kernel sizes. While both adaptations substantially decrease the total number of MACs required for inference, sustaining low latency necessitates preserving hardware-efficiency. Our experiments across diverse CPU devices confirm that these adaptations successfully retain high hardware-efficiency on CPUs. Based on these insights, we introduce CPUBone, a new family of vision backbone models optimized for CPU-based inference. CPUBone achieves state-of-the-art Speed-Accuracy Trade-offs (SATs) across a wide range of CPU devices and effectively transfers its efficiency to downstream tasks such as object detection and semantic segmentation. Models and code are available at https://github.com/altair199797/CPUBone.

This paper studies a class of multivariate Kantorovich-kernel neural network operators, including the deep Kantorovich-type neural network operators studied by Sharma and Singh. We prove density results, establish quantitative convergence estimates, derive Voronovskaya-type theorems, analyze the limits of partial differential equations for deep composite operators, prove Korovkin-type theorems, and propose inversion theorems. This paper studies a class of multivariate Kantorovich-kernel neural network operators, including the deep Kantorovich-type neural network operators studied by Sharma and Singh. We prove density results, establish quantitative convergence estimates, derive Voronovskaya-type theorems, analyze the limits of partial differential equations for deep composite operators, prove Korovkin-type theorems, and propose inversion theorems. Furthermore, this paper discusses the connection between neural network architectures and the classical positive operators proposed by Chui, Hsu, He, Lorentz, and Korovkin.

Uncertainty quantification is essential for deploying machine learning models in high-stakes domains such as scientific discovery and healthcare. Conformal Prediction (CP) provides finite-sample coverage guarantees under exchangeability, an assumption often violated in practice due to distribution shift. Under covariate shift, restoring validity requires importance weighting, yet accurate density-ratio estimation becomes unstable when training and test distributions exhibit limited support overlap. We propose KMM-CP, a conformal prediction framework based on Kernel Mean Matching (KMM) for covariate-shift correction. We show that KMM directly controls the bias-variance components governing conformal coverage error by minimizing RKHS moment discrepancy under explicit weight constraints, and establish asymptotic coverage guarantees under mild conditions. We then introduce a selective extension that identifies regions of reliable support overlap and restricts conformal correction to this subset, further improving stability in low-overlap regimes. Experiments on molecular property prediction benchmarks with realistic distribution shifts show that KMM-CP reduces coverage gap by over 50% compared to existing approaches. The code is available at https://github.com/siddharthal/KMM-CP.

Extended-thinking models expose a second text-generation channel ("thinking tokens") alongside the user-visible answer. This study examines 12 open-weight reasoning models on MMLU and GPQA questions paired with misleading hints. Among the 10,506 cases where models actually followed the hint (choosing the hint's target over the ground truth), each case is classified by whether the model acknowledges the hint in its thinking tokens, its answer text, both, or neither. In 55.4% of these cases the model's thinking tokens contain hint-related keywords that the visible answer omits entirely, a pattern termed *thinking-answer divergence*. The reverse (answer-only acknowledgment) is near-zero (0.5%), confirming that the asymmetry is directional. Hint type shapes the pattern sharply: sycophancy is the most *transparent* hint, with 58.8% of sycophancy-influenced cases acknowledging the professor's authority in both channels, while consistency (72.2%) and unethical (62.7%) hints are dominated by thinking-only acknowledgment. Models also vary widely, from near-total divergence (Step-3.5-Flash: 94.7%) to relative transparency (Qwen3.5-27B: 19.6%). These results show that answer-text-only monitoring misses more than half of all hint-influenced reasoning and that thinking-token access, while necessary, still leaves 11.8% of cases with no verbalized acknowledgment in either channel.

Comparison and evaluation of graph-based representations of sentence meaning is a challenge because competing representations of the same sentence may have different number of nodes, and it is not obvious which nodes should be compared to each other. Existing approaches favor node mapping that maximizes $F_1$ score over node relations and attributes, regardless whether the similarity is intentional or accidental; consequently, the identified mismatches in values of node attributes are not useful for any detailed error analysis. We propose a node-matching algorithm that allows comparison of multiple Uniform Meaning Representations (UMR) of one sentence and that takes advantage of node-word alignments, inherently available in UMR. We compare it with previously used approaches, in particular smatch (the de-facto standard in AMR evaluation), and argue that sensitivity to word alignment makes the comparison of meaning representations more intuitive and interpretable, while avoiding the NP-hard search problem inherent in smatch. A script implementing the method is freely available.

The Triplet Margin Ranking Loss is one of the most widely used loss functions in Siamese Networks for solving Distance Metric Learning (DML) problems. This loss function depends on a margin parameter μ, which defines the minimum distance that should separate positive and negative pairs during training. In this work, we show that, during training, the effective margin of many triplets often exceeds the predefined value of μ, provided that a sufficient number of triplets violating this margin is observed. This behavior indicates that fixing the margin throughout training may limit the learning process. Based on this observation, we propose a margin scheduler that adjusts the value of μ according to the proportion of easy triplets observed at each epoch, with the goal of maintaining training difficulty over time. We show that the proposed strategy leads to improved performance when compared to both a constant margin and a monotonically increasing margin scheme. Experimental results on four different datasets show consistent gains in verification performance.

The attention mechanism has been the core component in modern transformer architectures. However, the computation of standard full attention scales quadratically with the sequence length, serving as a major bottleneck in long-context language modeling. Sliding window attention restricts the context length for better efficiency at the cost of narrower receptive fields. While existing efforts attempt to take the benefits from both sides by building hybrid models, they often resort to static, heuristically designed alternating patterns that limit efficient allocation of computation in various scenarios. In this paper, we propose Switch Attention (SwiAttn), a novel hybrid transformer that enables dynamic and fine-grained routing between full attention and sliding window attention. For each token at each transformer layer, SwiAttn dynamically routes the computation to either a full-attention branch for global information aggregation or a sliding-window branch for efficient local pattern matching. An adaptive regularization objective is designed to encourage the model towards efficiency. Moreover, we adopt continual pretraining to optimize the model, transferring the full attention architecture to the hybrid one. Extensive experiments are conducted on twenty-three benchmark datasets across both regular (4K) and long (32K) context lengths, demonstrating the effectiveness of the proposed method.

Generative modeling has become a central paradigm in protein research, extending machine learning beyond structure prediction toward sequence design, backbone generation, inverse folding, and biomolecular interaction modeling. However, the literature remains fragmented across representations, model classes, and task formulations, making it difficult to compare methods or identify appropriate evaluation standards. This survey provides a systematic synthesis of generative AI in protein research, organized around (i) foundational representations spanning sequence, geometric, and multimodal encodings; (ii) generative architectures including $\mathrm{SE}(3)$-equivariant diffusion, flow matching, and hybrid predictor-generator systems; and (iii) task settings from structure prediction and de novo design to protein-ligand and protein-protein interactions. Beyond cataloging methods, we compare assumptions, conditioning mechanisms, and controllability, and we synthesize evaluation best practices that emphasize leakage-aware splits, physical validity checks, and function-oriented benchmarks. We conclude with critical open challenges: modeling conformational dynamics and intrinsically disordered regions, scaling to large assemblies while maintaining efficiency, and developing robust safety frameworks for dual-use biosecurity risks. By unifying architectural advances with practical evaluation standards and responsible development considerations, this survey aims to accelerate the transition from predictive modeling to reliable, function-driven protein engineering.

The generation of texts using Large Language Models (LLMs) is inherently uncertain, with sources of uncertainty being not only the generation of texts, but also the prompt used and the downstream interpretation. Within this work, we provide a formal framework for the measurement of uncertainty that takes these different aspects into account. Our framework models prompting, generation, and interpretation as interconnected autoregressive processes that can be combined into a single sampling tree. We introduce filters and objective functions to describe how different aspects of uncertainty can be expressed over the sampling tree and demonstrate how to express existing approaches towards uncertainty through these functions. With our framework we show not only how different methods are formally related and can be reduced to a common core, but also point out additional aspects of uncertainty that have not yet been studied.

Designing quantum algorithms is a complex and counterintuitive task, making it an ideal candidate for AI-driven algorithm discovery. To this end, we employ the Hive, an AI platform for program synthesis, which utilises large language models to drive a highly distributed evolutionary process for discovering new algorithms. We focus on the ground state problem in quantum chemistry, and discover efficient quantum heuristic algorithms that solve it for molecules LiH, H2O, and F2 while exhibiting significant reductions in quantum resources relative to state-of-the-art near-term quantum algorithms. Further, we perform an interpretability study on the discovered algorithms and identify the key functions responsible for the efficiency gains. Finally, we benchmark the Hive-discovered circuits on the Quantinuum System Model H2 quantum computer and identify minimum system requirements for chemical precision. We envision that this novel approach to quantum algorithm discovery applies to other domains beyond chemistry, as well as to designing quantum algorithms for fault-tolerant quantum computers.

Attention Deficit Hyperactivity Disorder (ADHD) is a highly prevalent neurodevelopmental condition; however, its neurobiological diagnosis remains challenging due to the lack of reliable imaging-based biomarkers, particularly anatomical markers. Structural MRI (sMRI) provides a non-invasive modality for investigating brain alterations associated with ADHD; nevertheless, most deep learning approaches function as black-box systems, limiting clinical trust and interpretability. In this work, we propose DuSCN-FusionNet, an interpretable sMRI-based framework for ADHD classification that leverages dual-channel Structural Covariance Networks (SCNs) to capture inter-regional morphological relationships. ROI-wise mean intensity and intra-regional variability descriptors are used to construct intensity-based and heterogeneity-based SCNs, which are processed through an SCN-CNN encoder. In parallel, auxiliary ROI-wise variability features and global statistical descriptors are integrated via late-stage fusion to enhance performance. The model is evaluated using stratified 10-fold cross-validation with a 5-seed ensemble strategy, achieving a mean balanced accuracy of 80.59% and an AUC of 0.778 on the Peking University site of the ADHD-200 dataset. DuSCN-FusionNet further achieves precision, recall, and F1-scores of 81.66%, 80.59%, and 80.27%, respectively. Moreover, Grad-CAM is adapted to the SCN domain to derive ROI-level importance scores, enabling the identification of structurally relevant brain regions as potential biomarkers.

Accurate uncertainty quantification is crucial for making reliable decisions in various supervised learning scenarios, particularly when dealing with complex, multimodal data such as images and text. Current approaches often face notable limitations, including rigid assumptions and limited generalizability, constraining their effectiveness across diverse supervised learning tasks. To overcome these limitations, we introduce Generative Score Inference (GSI), a flexible inference framework capable of constructing statistically valid and informative prediction and confidence sets across a wide range of multimodal learning problems. GSI utilizes synthetic samples generated by deep generative models to approximate conditional score distributions, facilitating precise uncertainty quantification without imposing restrictive assumptions about the data or tasks. We empirically validate GSI's capabilities through two representative scenarios: hallucination detection in large language models and uncertainty estimation in image captioning. Our method achieves state-of-the-art performance in hallucination detection and robust predictive uncertainty in image captioning, and its performance is positively influenced by the quality of the underlying generative model. These findings underscore the potential of GSI as a versatile inference framework, significantly enhancing uncertainty quantification and trustworthiness in multimodal learning.

Multimodal Large Language Models (MLLMs) achieve strong multimodal reasoning performance, yet we identify a recurring failure mode in long-form generation: as outputs grow longer, models progressively drift away from image evidence and fall back on textual priors, resulting in ungrounded reasoning and hallucinations. Interestingly, Based on attention analysis, we find that MLLMs have a latent capability for late-stage visual verification that is present but not consistently activated. Motivated by this observation, we propose Visual Re-Examination (VRE), a self-evolving training framework that enables MLLMs to autonomously perform visual introspection during reasoning without additional visual inputs. Rather than distilling visual capabilities from a stronger teacher, VRE promotes iterative self-improvement by leveraging the model itself to generate reflection traces, making visual information actionable through information gain. Extensive experiments across diverse multimodal benchmarks demonstrate that VRE consistently improves reasoning accuracy and perceptual reliability, while substantially reducing hallucinations, especially in long-chain settings. Code is available at https://github.com/Xiaobu-USTC/VRE.

The complex Gaussian distribution has been widely used as a fundamental spectral and noise model in signal processing and communication. However, its Gaussian structure often limits its ability to represent the diverse amplitude characteristics observed in individual source signals. On the other hand, many existing non-Gaussian amplitude distributions derived from hyperspherical models achieve good empirical fit due to their power-law structures, while they do not explicitly account for the complex-plane geometry inherent in complex-valued observations. In this paper, we propose a new probabilistic model for complex-valued random variables, which can be interpreted as a power-weighted noncentral complex Gaussian distribution. Unlike conventional hyperspherical amplitude models, the proposed model is formulated directly on the complex plane and preserves the geometric structure of complex-valued observations while retaining a higher-dimensional interpretation. The model introduces a nonlinear phase diffusion through a single shape parameter, enabling continuous control of the distributional geometry from arc-shaped diffusion along the phase direction to concentration of probability mass toward the origin. We formulate the proposed distribution and analyze the statistical properties of the induced amplitude distribution. The derived amplitude and power distributions provide a unified framework encompassing several widely used distributions in signal modeling, including the Rice, Nakagami, and gamma distributions. Experimental results on speech power spectra demonstrate that the proposed model consistently outperforms conventional distributions in terms of log-likelihood.

We propose an Expected Free Energy-based acquisition function for Bayesian optimization to solve the joint learning and optimization problem, i.e., optimize and learn the underlying function simultaneously. We show that, under specific assumptions, Expected Free Energy reduces to Upper Confidence Bound, Lower Confidence Bound, and Expected Information Gain. We prove that Expected Free Energy has unbiased convergence guarantees for concave functions. Using the results from these derivations, we introduce a curvature-aware update law for Expected Free Energy and show its proof of concept using a system identification problem on a Van der Pol oscillator. Through rigorous simulation experiments, we show that our adaptive Expected Free Energy-based acquisition function outperforms state-of-the-art acquisition functions with the least final simple regret and error in learning the Gaussian process.

Repository-level code agents have shown strong promise in real-world feature addition tasks, making reliable evaluation of their capabilities increasingly important. However, existing benchmarks primarily evaluate these agents as black boxes based on final test correctness, providing limited insight into how they reason and where failures arise. To address this limitation, we introduce RACE-bench, a reasoning-augmented benchmark for evaluating code agents on repository-level feature addition tasks. RACE-bench contains 528 real-world feature addition instances from 12 open-source repositories. Each instance is paired with executable patch verification and structured intermediate reasoning ground truth covering issue understanding, file localization, implementation tasks, and step decomposition. Based on this design, we introduce a dual-track evaluation framework that jointly measures patch correctness and intermediate reasoning quality. We evaluate three representative repository-level code agents on RACE-bench. On the full benchmark, Resolved Rates range from 29% to 70% across different agents. Our reasoning-level analysis further shows that while current agents perform well at understanding high-level intent, their performance degrades substantially when translating intent into concrete implementation steps. We also find that apply-success but test-fail cases exhibit lower reasoning recall (35.7% decrease) and higher over-prediction (94.1% increase) compared to successful cases. These findings highlight the importance of evaluating repository-level code agents beyond final patch correctness by examining the quality of their reasoning processes.

Legal reasoning requires not only the application of legal rules but also an understanding of the context in which those rules operate. However, existing legal benchmarks primarily evaluate rule application under the assumption of fixed norms, and thus fail to capture situations where legal judgments shift or where multiple norms interact. In this work, we propose CALRK-Bench, a context-aware legal reasoning benchmark based on the legal system in Korean. CALRK-Bench evaluates whether models can identify the temporal validity of legal norms, determine whether sufficient legal information is available for a given case, and understand the reasons behind shifts in legal judgments. The dataset is constructed from legal precedents and legal consultation records, and is validated by legal experts. Experimental results show that even recent large language models consistently exhibit low performance on these three tasks. CALRK-Bench provides a new stress test for evaluating context-aware legal reasoning rather than simple memorization of legal knowledge. Our code is available at https://github.com/jhCOR/CALRKBench.

Supervised fine-tuning (SFT) on visual instruction data often improves perceptual capabilities in vision-language models (VLMs) while degrading reasoning performance, creating a persistent reasoning tax during post-training. We investigate whether this degradation is related to disrupted access to depth-wise representations, and find that even fixed cross-depth aggregation substantially restores reasoning, suggesting that preserved cross-depth access is an important missing factor in VLM fine-tuning. Building on this observation, we propose Input-Adaptive Depth Aggregation (IADA), a lightweight mechanism that makes cross-depth retrieval input-adaptive, modality-aware, and efficiently parameterized through a low-rank bottleneck. On Qwen3-VL-2B, IADA improves the average reasoning score by 9.5 points and the average perception score by $3.3$ points over LoRA-only fine-tuning with only 0.14M additional parameters, with the strongest gains appearing in parameter-efficient low-rank settings.

The rapid integration of Large Language Models (LLMs) into software engineering practice is reshaping how software testing activities are performed. LLMs are increasingly used to support software testing. Consequently, software testing education must evolve to prepare students for this new paradigm. However, while students have already begun to use LLMs in an ad hoc manner for testing tasks, there is limited empirical understanding of how such usage influences their testing behaviors, judgment, and learning outcomes. It is necessary to conduct a systematic investigation into how students learn to evaluate, control, and refine LLM-assisted testing results. This paper presents a mixed-methods, two-phase exploratory study on human-LLM collaboration in software testing education. In Phase I, we analyze classroom learning artifacts and interaction records from 15 students, together with a large-scale survey conducted in a national software testing competition (337 valid responses), to identify recurring prompt-related difficulties across testing tasks. The results reveal systematic interaction breakdowns, including missing contextual information, insufficient constraints, rigid one-shot prompting, and limited strategy-driven iteration, with automated test script generation emerging as a particularly heterogeneous and effort-intensive interaction context. Building on these findings, Phase II conducts an illustrative classroom practice that operationalizes the observed breakdowns into a lightweight, stage-aware prompt scaffold for test script generation, guiding students to explicitly articulate execution-relevant information such as environmental assumptions, interaction grounding, synchronization, and validation intent, and reporting descriptive shifts in students' testing-related articulation when interacting with LLMs.

In this paper we develop a graph-learning algorithm, MED-MAGMA, to fit multi-axis (Kronecker-sum-structured) models corrupted by multiplicative noise. This type of noise is natural in many application domains, such as that of single-cell RNA sequencing, in which it naturally captures technical biases of RNA sequencing platforms. Our work is evaluated against prior work on each and every public dataset in the Single Cell Expression Atlas under a certain size, demonstrating that our methodology learns networks with better local and global structure. MED-MAGMA is made available as a Python package (MED-MAGMA).

Most existing approaches to AI in pharmacy collapse three epistemologically distinct operations into a single technical layer: document preservation, semantic interpretation, and contextual presentation. This conflation is a root cause of recurring fragilities including loss of provenance, interpretive opacity, alert fatigue, and erosion of accountability. This paper proposes the PATOS--Lector--PRISMA (PLP) infrastructure as a normative information architecture for responsible pharmaceutical knowledge management. PATOS preserves regulatory documents with explicit versioning and provenance; Lector implements machine-assisted reading with human curation, producing typed assertions anchored to primary sources; PRISMA delivers contextual presentation through the RPDA framework (Regulatory, Prescription, Dispensing, Administration), refracting the same informational core into distinct professional views. The architecture introduces the Evidence Pack as a formal unit of accountable assertion (versioned, traceable, epistemically bounded, and curatorially validated), with assertions typified by illocutionary force. A worked example traces dipyrone monohydrate across all three layers using real system data. Developed and validated in Brazil's regulatory context, the architecture is grounded in an operational implementation comprising over 16,000 official documents and 38 curated Evidence Packs spanning five reference medications. The proposal is demonstrated as complementary to operational decision support systems, providing infrastructural conditions that current systems lack: documentary anchoring, interpretive transparency, and institutional accountability.

As spatial intelligence becomes an increasingly important capability for foundation models, it remains unclear whether large language models' (LLMs) performance on spatial reasoning benchmarks reflects structured internal spatial representations or reliance on linguistic heuristics. We address this question from a mechanistic perspective by examining how spatial information is internally represented and used. Drawing on computational theories of human spatial cognition, we decompose spatial reasoning into three primitives, relational composition, representational transformation, and stateful spatial updating, and design controlled task families for each. We evaluate multilingual LLMs in English, Chinese, and Arabic under single pass inference, and analyze internal representations using linear probing, sparse autoencoder based feature analysis, and causal interventions. We find that task relevant spatial information is encoded in intermediate layers and can causally influence behavior, but these representations are transient, fragmented across task families, and weakly integrated into final predictions. Cross linguistic analysis further reveals mechanistic degeneracy, where similar behavioral performance arises from distinct internal pathways. Overall, our results suggest that current LLMs exhibit limited and context dependent spatial representations rather than robust, general purpose spatial reasoning, highlighting the need for mechanistic evaluation beyond benchmark accuracy.

We develop a tensor-network surrogate for option pricing, targeting large-scale portfolio revaluation problems arising in market risk management (e.g., VaR and Expected Shortfall computations). The method involves representing high-dimensional price surfaces in tensor-train (TT) form using TT-cross approximation, constructing the surrogate directly from black-box price evaluations without materializing the full training tensor. For inference, we use a Laplacian kernel and derive TT representations of the kernel matrix and its closed-form inverse in the noise-free setting, enabling TT-based Gaussian process regression without dense matrix factorization or iterative linear solves. We found that hyperparameter optimization consistently favors a large kernel length-scale and show that in this regime the GPR predictor reduces to multilinear interpolation for off-grid inputs; we also derive a low-rank TT representation for this limit. We evaluate the approach on five-asset basket options over an eight dimensional parameter space (asset spot levels, strike, interest rate, and time to maturity). For European geometric basket puts, the tensor surrogate achieves lower test error at shorter training times than standard GPR by scaling to substantially larger effective training sets. For American arithmetic basket puts trained on LSMC data, the surrogate exhibits more favorable scaling with training-set size while providing millisecond-level evaluation per query, with overall runtime dominated by data generation.

Model merging combines independently fine-tuned checkpoints without joint multi-task training. In the era of foundation-model, fine-tuning with Low-Rank Adaptation (LoRA) is prevalent, making LoRA merging a promising target. Existing approaches can work in homogeneous settings where all target tasks are classification but often fail when tasks span classification and regression. Approaches using entropy-based surrogates do not apply to regression and are costly for large language models due to long token sequences. We introduce Null-Space Compression (NSC) Merging, a label-free, output-agnostic method that sets merge weights from adapter geometry. Our key observation is that during LoRA finetuning the down-projection factor $A$ in $ΔW = BA$ compresses its null space, and the compression correlates with performance. NSC uses this as an optimization signal for merging that can generalize across classification, regression, and sequence generation. NSC achieves state-of-the-art performance across twenty heterogeneous vision tasks with balanced gains where prior methods overfit subsets of tasks. It also outperforms baselines on six NLI benchmarks and on vision-language evaluations for VQA and image captioning, demonstrating scalability and effectiveness.

As multimodal models like CLIP become integral to downstream systems, the need to remove sensitive information is critical. However, machine unlearning for contrastively-trained encoders remains underexplored, and existing evaluations fail to diagnose fine-grained, association-level forgetting. We introduce SALMUBench (Sensitive Association-Level Multimodal Unlearning), a benchmark built upon a synthetic dataset of 60K persona-attribute associations and two foundational models: a Compromised model polluted with this data, and a Clean model without it. To isolate unlearning effects, both are trained from scratch on the same 400M-pair retain base, with the Compromised model additionally trained on the sensitive set. We propose a novel evaluation protocol with structured holdout sets (holdout identity, holdout association) to precisely measure unlearning efficacy and collateral damage. Our benchmark reveals that while utility-efficient deletion is feasible, current methods exhibit distinct failure modes: they either fail to forget effectively or over-generalize by erasing more than intended. SALMUBench sets a new standard for comprehensive unlearning evaluation, and we publicly release our dataset, models, evaluation scripts, and leaderboards to foster future research.

We propose a semi-structured discrete-time multi-state model to analyse mortgage delinquency transitions. This model combines an easy-to-understand structured additive predictor, which includes linear effects and smooth functions of time and covariates, with a flexible neural network component that captures complex nonlinearities and higher-order interactions. To ensure identifiability when covariates are present in both components, we orthogonalise the unstructured part relative to the structured design. For discrete-time competing transitions, we derive exact transformations that map binary logistic models to valid competing transition probabilities, avoiding the need for continuous-time approximations. In simulations, our framework effectively recovers structured baseline and covariate effects while using the neural component to detect interaction patterns. We demonstrate the method using the Freddie Mac Single-Family Loan-Level Dataset, employing an out-of-time test design. Compared with a structured generalised additive benchmark, the semi-structured model provides modest but consistent gains in discrimination across the earliest prediction spans, while maintaining similar Brier scores. Adding macroeconomic indicators provides limited incremental benefit in this out-of-time evaluation and does not materially change the estimated borrower-, loan-, or duration-driven effects. Overall, semi-structured multi-state modelling offers a practical compromise between transparent effect estimates and flexible pattern learning, with potential applications beyond credit-transition forecasting.

Attention Deficit Hyperactivity Disorder (ADHD) is a prevalent neurodevelopmental disorder whose neuroimaging-based diagnosis remains challenging due to complex time-varying disruptions in brain connectivity. Functional MRI (fMRI) provides a powerful non-invasive modality for identifying functional alterations. Existing deep learning (DL) studies employ diverse neuroimaging features; however, static functional connectivity remains widely used, whereas dynamic connectivity modeling is comparatively underexplored. Moreover, many DL models lack interpretability. In this work, we propose D-GATNet, an interpretable temporal graph-based framework for automated ADHD classification using dynamic functional connectivity (dFC). Sliding-window Pearson correlation constructs sequences of functional brain graphs with regions of interest as nodes and connectivity strengths as edges. Spatial dependencies are learned via a multi-layer Graph Attention Network, while temporal dynamics are modeled using 1D convolution followed by temporal attention. Interpretability is achieved through graph attention weights revealing dominant ROI interactions, ROI importance scores identifying influential regions, and temporal attention emphasizing informative connectivity segments. Experiments on the Peking University site of the ADHD-200 dataset using stratified 10-fold cross-validation with a 5-seed ensemble achieved 85.18% +_5.64 balanced accuracy and 0.881 AUC, outperforming state-of-the-art methods. Attention analysis reveals cerebellar and default mode network disruptions, indicating potential neuroimaging biomarkers.

Merging multiple Low-Rank Adaptation (LoRA) modules is promising for constructing general-purpose systems, yet challenging because LoRA update directions span different subspaces and contribute unevenly. When merged naively, such mismatches can weaken the directions most critical to certain task losses while overemphasizing relatively less important ones, ultimately reducing the model's ability to represent all tasks faithfully. We revisit this problem through two perspectives: subspace coverage, which captures how broadly LoRA directions cover diverse representational directions, and anisotropy, which reflects the imbalance of influence across those directions. We propose TARA-Merging (Task-Rank Anisotropy Alignment), which aligns merging weights using a preference-weighted cross-entropy pseudo-loss while preserving task-relevant LoRA subspaces. This ensures broad subspace coverage and mitigates anisotropy via direction-wise reweighting. Across eight vision and six NLI benchmarks, TARA-Merging consistently outperforms vanilla and LoRA-aware baselines, demonstrating strong robustness and generalization, and highlighting the importance of addressing both subspace coverage and anisotropy in LoRA merging.

Syllable-level units offer compact and linguistically meaningful representations for spoken language modeling and unsupervised word discovery, but research on syllabification remains fragmented across disparate implementations, datasets, and evaluation protocols. We introduce findsylls, a modular, language-agnostic toolkit that unifies classical syllable detectors and end-to-end syllabifiers under a common interface for syllable segmentation, embedding extraction, and multi-granular evaluation. The toolkit implements and standardizes widely used methods (e.g., Sylber, VG-HuBERT) and allows their components to be recombined, enabling controlled comparisons of representations, algorithms, and token rates. We demonstrate findsylls on English and Spanish corpora and on new hand-annotated data from Kono, an underdocumented Central Mande language, illustrating how a single framework can support reproducible syllable-level experiments across both high-resource and under-resourced settings.

Generative video models achieve high visual fidelity but often violate basic physical principles, limiting reliability in real-world settings. Prior attempts to inject physics rely on conditioning: frame-level signals are domain-specific and short-horizon, while global text prompts are coarse and noisy, missing fine-grained dynamics. We present PhysVid, a physics-aware local conditioning scheme that operates over temporally contiguous chunks of frames. Each chunk is annotated with physics-grounded descriptions of states, interactions, and constraints, which are fused with the global prompt via chunk-aware cross-attention during training. At inference, we introduce negative physics prompts (descriptions of locally relevant law violations) to steer generation away from implausible trajectories. On VideoPhy, PhysVid improves physical commonsense scores by $\approx 33\%$ over baseline video generators, and by up to $\approx 8\%$ on VideoPhy2. These results show that local, physics-aware guidance substantially increases physical plausibility in generative video and marks a step toward physics-grounded video models.

The availability of generative Artificial Intelligence (AI) tools such as ChatGPT or GitHub Copilot is reshaping the way in which software is developed, evolved, and maintained. Oftentimes, developers leave traces of such an usage in software artifacts. This allows not only to understand how AI is used in software development, but also to let others be aware how such software artifacts were created, e.g., for licensing or trustworthiness purposes. This paper-building upon our preliminary work presented at MSR 2024-aims at qualitatively investigating on the self-admitted use of two very popular generative AI tools - ChatGPT and GitHub Copilot - in software development. To this aim, we mined GitHub for such traces, by looking at commits, issues and pull requests (PRs). Then, through a manual coding, we create a taxonomy of 64 different ChatGPT and GitHub Copilot usage tasks, grouped into 7 categories. By repeating our previous analysis two years after and by extending it to GitHub Copilot, we show how the usage avenues have been expanded, the extent to which developers perceived such a generative AI usage useful, and whether some concerns occurring more than one year ago are no longer present. The taxonomy of tasks we derived from such a qualitative study provided (i) developers with valuable insights into how generative AI can be integrated into their workflows, and (ii) researchers with a clear overview of tasks that developers perceive as well-suited for automation.

Smart contracts govern billions of dollars in decentralized finance (DeFi), yet automated vulnerability detection remains challenging because many vulnerabilities are tightly coupled with project-specific business logic. We observe that recurring vulnerabilities across diverse DeFi business models often share the same underlying economic mechanisms, which we term DeFi semantics, and that capturing these shared abstractions can enable more systematic auditing. Building on this insight, we propose Knowdit, a knowledge-driven, agentic framework for smart contract vulnerability detection. Knowdit first constructs an auditing knowledge graph from historical human audit reports, linking fine-grained DeFi semantics with recurring vulnerability patterns. Given a new project, a multi-agent framework leverages this knowledge through an iterative loop of specification generation, harness synthesis, fuzz execution, and finding reflection, driven by a shared working memory for continuous refinement. We evaluate Knowdit on 12 recent Code4rena projects with 75 ground-truth vulnerabilities. Knowdit detects all 14 high-severity and 77\% of medium-severity vulnerabilities with only 2 false positives, significantly outperforming all baselines. Applied to six real-world projects, Knowdit further discovers 12 high- and 10 medium-severity previously unknown vulnerabilities, proving its outstanding performance.

Large vision-language models have endowed GUI agents with strong general capabilities for interface understanding and interaction. However, due to insufficient exposure to domain-specific software operation data during training, these agents exhibit significant domain bias - they lack familiarity with the specific operation workflows (planning) and UI element layouts (grounding) of particular applications, limiting their real-world task performance. In this paper, we present GUIDE (GUI Unbiasing via Instructional-Video Driven Expertise), a training-free, plug-and-play framework that resolves GUI agent domain bias by autonomously acquiring domain-specific expertise from web tutorial videos through a retrieval-augmented automated annotation pipeline. GUIDE introduces two key innovations. First, a subtitle-driven Video-RAG pipeline unlocks video semantics through subtitle analysis, performing progressive three-stage retrieval - domain classification, topic extraction, and relevance matching - to identify task-relevant tutorial videos. Second, a fully automated annotation pipeline built on an inverse dynamics paradigm feeds consecutive keyframes enhanced with UI element detection into VLMs, inferring the required planning and grounding knowledge that are injected into the agent's corresponding modules to address both manifestations of domain bias. Extensive experiments on OSWorld demonstrate GUIDE's generality as a plug-and-play component for both multi-agent systems and single-model agents. It consistently yields over 5% improvements and reduces execution steps - without modifying any model parameters or architecture - validating GUIDE as an architecture-agnostic enhancement to bridge GUI agent domain bias.

Optimal dispatch of energy storage systems (ESSs) in distribution networks involves jointly improving operating economy and voltage security under time-varying conditions and possible topology changes. To support fast online decision making, we develop a topology-aware Reinforcement Learning architecture based on Twin Delayed Deep Deterministic Policy Gradient (TD3), which integrates graph neural networks (GNNs) as graph feature encoders for ESS dispatch. We conduct a systematic investigation of three GNN variants: graph convolutional networks (GCNs), topology adaptive graph convolutional networks (TAGConv), and graph attention networks (GATs) on the 34-bus and 69-bus systems, and evaluate robustness under multiple topology reconfiguration cases as well as cross-system transfer between networks with different system sizes. Results show that GNN-based controllers consistently reduce the number and magnitude of voltage violations, with clearer benefits on the 69-bus system and under reconfiguration; on the 69-bus system, TD3-GCN and TD3-TAGConv also achieve lower saved cost relative to the NLP benchmark than the NN baseline. We also highlight that transfer gains are case-dependent, and zero-shot transfer between fundamentally different systems results in notable performance degradation and increased voltage magnitude violations. This work is available at: https://github.com/ShuyiGao/GNNs_RL_ESSs and https://github.com/distributionnetworksTUDelft/GNNs_RL_ESSs.

Image-to-point cloud registration methods typically follow a coarse-to-fine pipeline, extracting patch-level correspondences and refining them into dense pixel-to-point matches. However, in scenes with repetitive patterns, images often lack sufficient 3D structural cues and alignment with point clouds, leading to incorrect matches. Moreover, prior methods usually overlook structural consistency, limiting the full exploitation of correspondences. To address these issues, we propose two novel modules: the Local Geometry Enhancement (LGE) module and the Graph Distribution Consistency (GDC) module. LGE enhances both image and point cloud features with normal vectors, injecting geometric structure into image features to reduce mismatches. GDC constructs a graph from matched points to update features and explicitly constrain similarity distributions. Extensive experiments and ablations on two benchmarks, RGB-D Scenes v2 and 7-Scenes, demonstrate that our approach achieves state-of-the-art performance in image-to-point cloud registration.

Contrastive learning produces coherent semantic feature embeddings by encouraging positive samples to cluster closely while separating negative samples. However, existing contrastive learning methods lack principled guarantees on coverage within the semantic feature space. We extend conformal prediction to this setting by introducing minimum-volume covering sets equipped with learnable generalized multi-norm constraints. We propose a method that constructs conformal sets guaranteeing user-specified coverage of positive samples while maximizing negative sample exclusion. We establish theoretically that volume minimization serves as a proxy for negative exclusion, enabling our approach to operate effectively even when negative pairs are unavailable. The positive inclusion guarantee inherits the distribution-free coverage property of conformal prediction, while negative exclusion is maximized through learned set geometry optimized on a held-out training split. Experiments on simulated and real-world image datasets demonstrate improved inclusion-exclusion trade-offs compared to standard distance-based conformal baselines.

Open-vocabulary 3D semantic segmentation aims to segment arbitrary categories beyond the training set. Existing methods predominantly rely on distilling knowledge from 2D open-vocabulary models. However, aligning 3D features to the 2D representation space restricts intrinsic 3D geometric learning and inherits errors from 2D predictions. To address these limitations, we propose GeoGuide, a novel framework that leverages pretrained 3D models to integrate hierarchical geometry-semantic consistency for open-vocabulary 3D segmentation. Specifically, we introduce an Uncertainty-based Superpoint Distillation module to fuse geometric and semantic features for estimating per-point uncertainty, adaptively weighting 2D features within superpoints to suppress noise while preserving discriminative information to enhance local semantic consistency. Furthermore, our Instance-level Mask Reconstruction module leverages geometric priors to enforce semantic consistency within instances by reconstructing complete instance masks. Additionally, our Inter-Instance Relation Consistency module aligns geometric and semantic similarity matrices to calibrate cross-instance consistency for same-category objects, mitigating viewpoint-induced semantic drift. Extensive experiments on ScanNet v2, Matterport3D, and nuScenes demonstrate the superior performance of GeoGuide.

While Late Interaction models exhibit strong retrieval performance, many of their underlying dynamics remain understudied, potentially hiding performance bottlenecks. In this work, we focus on two topics in Late Interaction retrieval: a length bias that arises when using multi-vector scoring, and the similarity distribution beyond the best scores pooled by the MaxSim operator. We analyze these behaviors for state-of-the-art models on the NanoBEIR benchmark. Results show that while the theoretical length bias of causal Late Interaction models holds in practice, bi-directional models can also suffer from it in extreme cases. We also note that no significant similarity trend lies beyond the top-1 document token, validating that the MaxSim operator efficiently exploits the token-level similarity scores.

We present ARTA, a mixed-resolution coarse-to-fine vision transformer for efficient dense feature extraction. Unlike models that begin with dense high-resolution (fine) tokens, ARTA starts with low-resolution (coarse) tokens and uses a lightweight allocator to predict which regions require more fine tokens. The allocator iteratively predicts a semantic (class) boundary score and allocates additional tokens to patches above a low threshold, concentrating token density near boundaries while maintaining high sensitivity to weak boundary evidence. This targeted allocation encourages tokens to represent a single semantic class rather than a mixture of classes. Mixed-resolution attention enables interaction between coarse and fine tokens, focusing computation on semantically complex areas while avoiding redundant processing in homogeneous regions. Experiments demonstrate that ARTA achieves state-of-the-art results on ADE20K and COCO-Stuff with substantially fewer FLOPs, and delivers competitive performance on Cityscapes at markedly lower compute. For example, ARTA-Base attains 54.6 mIoU on ADE20K in the ~100M-parameter class while using fewer FLOPs and less memory than comparable backbones.

Hypertrophic cardiomyopathy (HCM) requires accurate risk stratification to inform decisions regarding ICD therapy and follow-up management. Current established models, such as the European Society of Cardiology (ESC) score, exhibit moderate discriminative performance. This study develops a robust, explainable machine learning (ML) risk score leveraging routinely collected echocardiographic, clinical, and medication data, typically contained within Electronic Health Records (EHRs), to predict a 5-year composite cardiovascular outcome in HCM patients. The model was trained and internally validated using a large cohort (N=1,201) from the SHARE registry (Florence Hospital) and externally validated on an independent cohort (N=382) from Rennes Hospital. The final Random Forest ensemble model achieved a high internal Area Under the Curve (AUC) of 0.85 +- 0.02, significantly outperforming the ESC score (0.56 +- 0.03). Critically, survival curve analysis on the external validation set showed superior risk separation for the ML score (Log-rank p = 8.62 x 10^(-4) compared to the ESC score (p = 0.0559). Furthermore, longitudinal analyses demonstrate that the proposed risk score remains stable over time in event-free patients. The model high interpretability and its capacity for longitudinal risk monitoring represent promising tools for the personalized clinical management of HCM.

Big data research in Indonesia is constrained by a fundamental fragmentation: relevant data is scattered across social media, news portals, e-commerce platforms, review sites, and academic databases, each with different formats, access methods, and noise characteristics. Researchers must independently build collection pipelines, clean heterogeneous data, and assemble separate analysis tools, a process that often overshadows the research itself. We present SocialX, a modular platform for multi-source big data research that integrates heterogeneous data collection, language-aware preprocessing, and pluggable analysis into a unified, source-agnostic pipeline. The platform separates concerns into three independent layers (collection, preprocessing, and analysis) connected by a lightweight job-coordination mechanism. This modularity allows each layer to grow independently: new data sources, preprocessing methods, or analysis tools can be added without modifying the existing pipeline. We describe the design principles that enable this extensibility, detail the preprocessing methodology that addresses challenges specific to Indonesian text across registers, and demonstrate the platform's utility through a walkthrough of a typical research workflow. SocialX is publicly accessible as a web-based platform at https://www.socialx.id.

AI accessibility tools have mostly been designed for individual use, helping one person overcome a specific functional barrier. But for many people with disabilities, complex tasks are accomplished through collaboration with others who bring complementary abilities, not solitary effort. We propose a three-layer framework, Channelling, Coordinating, and Co-Creating, that rethinks AI's role in ability-diverse collaboration: establishing shared informational ground across abilities, mediating workflows between collaborators with different abilities, and contributing as a bounded partner toward shared goals. Grounded in the Ability-Diverse Collaboration framework, grounding theory, and Carlile's 3T framework, it extends the ``agents as remote collaborators'' vision by centring the collaborative, interdependent ways people with disabilities already work.

Transformer-based reinforcement learning has emerged as a strong candidate for sequential control in residential energy management. In particular, the Decision Transformer can learn effective battery dispatch policies from historical data, thereby increasing photovoltaic self-consumption and reducing electricity costs. However, transformer models are typically too computationally demanding for deployment on resource-constrained residential controllers, where memory and latency constraints are critical. This paper investigates knowledge distillation to transfer the decision-making behaviour of high-capacity Decision Transformer policies to compact models that are more suitable for embedded deployment. Using the Ausgrid dataset, we train teacher models in an offline sequence-based Decision Transformer framework on heterogeneous multi-building data. We then distil smaller student models by matching the teachers' actions, thereby preserving control quality while reducing model size. Across a broad set of teacher-student configurations, distillation largely preserves control performance and even yields small improvements of up to 1%, while reducing the parameter count by up to 96%, the inference memory by up to 90%, and the inference time by up to 63%. Beyond these compression effects, comparable cost improvements are also observed when distilling into a student model of identical architectural capacity. Overall, our results show that knowledge distillation makes Decision Transformer control more applicable for residential energy management on resource-limited hardware.

Language endangerment poses a major challenge to linguistic diversity worldwide, and technological advances have opened new avenues for documentation and revitalization. Among these, automatic speech recognition (ASR) has shown increasing potential to assist in the transcription of endangered language data. This study focuses on Ikema, a severely endangered Ryukyuan language spoken in Okinawa, Japan, with approximately 1,300 remaining speakers, most of whom are over 60 years old. We present an ongoing effort to develop an ASR system for Ikema based on field recordings. Specifically, we (1) construct a {\totaldatasethours}-hour speech corpus from field recordings, (2) train an ASR model that achieves a character error rate as low as 15\%, and (3) evaluate the impact of ASR assistance on the efficiency of speech transcription. Our results demonstrate that ASR integration can substantially reduce transcription time and cognitive load, offering a practical pathway toward scalable, technology-supported documentation of endangered languages.

Standard LLM-based speech recognition systems typically process utterances in isolation, limiting their ability to leverage conversational context. In this work, we study whether multimodal context from prior turns improves LLM-based ASR and how to represent that context efficiently. We find that, after supervised multi-turn training, conversational context mainly helps with the recognition of contextual entities. However, conditioning on raw context is expensive because the prior-turn audio token sequence grows rapidly with conversation length. To address this, we propose Abstract Compression, which replaces the audio portion of prior turns with a fixed number of learned latent tokens while retaining corresponding transcripts explicitly. On both in-domain and out-of-domain test sets, the compressed model recovers part of the gains of raw-context conditioning with a smaller prior-turn audio footprint. We also provide targeted analyses of the compression setup and its trade-offs.

In a previous work (Elaarabi et al., 2025b), the Sequence Encoder for online dynamical system identification (Elaarabi et al., 2025a) and its combination with PINN (PINN-SE) were introduced and tested on both synthetic and real data case scenarios. The sequence encoder is able to effectively encode time series into feature vectors, which the PINN then uses to map to dynamical behavior, predicting system response under changes in parameters, ICs and BCs. Previously (Elaarabi et al., 2025b), the tests on real data were limited to simple 1D problems and only 1D time series inputs of the Sequence Encoder. In this work, the possibility of applying PINN-SE to a more realistic case is investigated: heating and early cooling of the thermo-stamping process, which is a critical stage in the forming process of continuous fiber reinforced composite materials with thermoplastic polymer. The possibility of extending the PINN-SE inputs to multimodal data, such as sequences of temporal 2D images and to scenarios involving variable geometries, is also explored. The results show that combining multiple encoders with the previously proposed method (Elaarabi et al., 2025b) is feasible, we also show that training the model on synthetic data generated based on experimental data can help the model to generalize well for real experimental data, unseen during the training phase.

Domain-driven design (DDD) is a powerful design technique for architecting complex software systems. This paper introduces a prompting framework that automates core DDD activities through structured large language model (LLM) interactions. We decompose DDD into five sequential steps: (1) establishing an ubiquitous language, (2) simulating event storming, (3) identifying bounded contexts, (4) designing aggregates, and (5) mapping to technical architecture. In a case study, we validated the prompting framework against real-world requirements from FTAPI's enterprise platform. While the first steps consistently generate valuable and usable artifacts, later steps show how minor errors or inaccuracies can propagate and accumulate. Overall, the framework excels as a collaborative sparring partner for building actionable documentation, such as glossaries and context maps, but not for full automation. This allows the experts to concentrate their discussion on the critical trade-offs. In our evaluation, Steps 1 to 3 worked well, but the accumulated errors rendered the artifacts generated from Steps 4 and 5 impractical. Our findings show that LLMs can enhance, but not replace, architectural expertise, offering a practical tool to reduce the effort and overhead of DDD while preserving human-centric decision-making.

While multilingual language models successfully transfer factual and syntactic knowledge across languages, it remains unclear whether they process culture-specific pragmatic registers, such as slang, as isolated language-specific memorizations or as unified, abstract concepts. We study this by probing the internal representations of Gemma-2-9B-IT using Sparse Autoencoders (SAEs) across three typologically diverse source languages: English, Hebrew, and Russian. To definitively isolate pragmatic register processing from trivial lexical sensitivity, we introduce a novel dataset in which every target term is polysemous, appearing in both literal and informal contexts. We find that while much of the informal-register signal is distributed across language-specific features, a small but highly robust cross-linguistic core consistently emerges. This shared core forms a geometrically coherent ``informal register subspace'' that sharpens in the model's deeper layers. Crucially, these shared representations are not merely correlational: activation steering with these features causally shifts output formality across all source languages and transfers zero-shot to six unseen languages spanning diverse language families and scripts. Together, these results provide the first mechanistic evidence that multilingual LLMs internalize informal register not just as surface-level heuristics, but as a portable, language-agnostic pragmatic abstraction.

We present GS-BrainText, a curated dataset of 8,511 brain radiology reports from the Generation Scotland cohort, of which 2,431 are annotated for 24 brain disease phenotypes. This multi-site dataset spans five Scottish NHS health boards and includes broad age representation (mean age 58, median age 53), making it uniquely valuable for developing and evaluating generalisable clinical natural language processing (NLP) algorithms and tools. Expert annotations were performed by a multidisciplinary clinical team using an annotation schema, with 10-100% double annotation per NHS health board and rigorous quality assurance. Benchmark evaluation using EdIE-R, an existing rule-based NLP system developed in conjunction with the annotation schema, revealed some performance variation across health boards (F1: 86.13-98.13), phenotypes (F1: 22.22-100) and age groups (F1: 87.01-98.13), highlighting critical challenges in generalisation of NLP tools. The GS-BrainText dataset addresses a significant gap in available UK clinical text resources and provides a valuable resource for the study of linguistic variation, diagnostic uncertainty expression and the impact of data characteristics on NLP system performance.

As Large Language Model (LLM) agents are increasingly deployed in open-ended domains like software engineering, they frequently encounter underspecified instructions that lack crucial context. While human developers naturally resolve underspecification by asking clarifying questions, current agents are largely optimized for autonomous execution. In this work, we systematically evaluate the clarification-seeking abilities of LLM agents on an underspecified variant of SWE-bench Verified. We propose an uncertainty-aware multi-agent scaffold that explicitly decouples underspecification detection from code execution. Our results demonstrate that this multi-agent system using OpenHands + Claude Sonnet 4.5 achieves a 69.40% task resolve rate, significantly outperforming a standard single-agent setup (61.20%) and closing the performance gap with agents operating on fully specified instructions. Furthermore, we find that the multi-agent system exhibits well-calibrated uncertainty, conserving queries on simple tasks while proactively seeking information on more complex issues. These findings indicate that current models can be turned into proactive collaborators, where agents independently recognize when to ask questions to elicit missing information in real-world, underspecified tasks.

Quantum Approximate Optimization Algorithm (QAOA) has emerged as a promising solution for combinatorial optimization problems using a hybrid quantum-classical framework. Among combinatorial optimization problems, the Maximum Cut (Max-Cut) problem is particularly important due to its broad applicability in various domains. While QAOA-based Max-Cut solvers have been developed, they primarily favor solution accuracy over execution efficiency, which significantly limits their practicality for large-scale problems. To address the limitation, we propose ParaQAOA, a parallel divide-and-conquer QAOA framework that leverages parallel computing hardware to efficiently solve large Max-Cut problems. ParaQAOA significantly reduces runtime by partitioning large problems into subproblems and solving them in parallel while preserving solution quality. This design not only scales to graphs with tens of thousands of vertices but also provides tunable control over accuracy-efficiency trade-offs, making ParaQAOA adaptable to diverse performance requirements. Experimental results demonstrate that ParaQAOA achieves up to 1,600x speedup over state-of-the-art methods on Max-Cut problems with 400 vertices while maintaining solution accuracy within 2% of the best-known solutions. Furthermore, ParaQAOA solves a 16,000-vertex instance in 19 minutes, compared to over 13.6 days required by the best-known approach. These findings establish ParaQAOA as a practical and scalable framework for large-scale Max-Cut problems under stringent time constraints.

Synchronous federated learning scales poorly due to the straggler effect. Asynchronous algorithms increase the update throughput by processing updates upon arrival, but they introduce two fundamental challenges: gradient staleness, which degrades convergence, and bias toward faster clients under heterogeneous data distributions. Although algorithms such as AsyncSGD and Generalized AsyncSGD mitigate this bias via client-side task queues, most existing analyses neglect the underlying queueing dynamics and lack closed-form characterizations of the update throughput and gradient staleness. To close this gap, we develop a stochastic queueing-network framework for Generalized AsyncSGD that jointly models random computation times at the clients and the central server, as well as random uplink and downlink communication delays. Leveraging product-form network theory, we derive a closed-form expression for the update throughput, alongside closed-form upper bounds for both the communication round complexity and the expected wall-clock time required to reach an $ε$-stationary point. These results formally characterize the trade-off between gradient staleness and wall-clock convergence speed. We further extend the framework to quantify energy consumption under stochastic timing, revealing an additional trade-off between convergence speed and energy efficiency. Building on these analytical results, we propose gradient-based optimization strategies to jointly optimize routing and concurrency. Experiments on EMNIST demonstrate reductions of 29%--46% in convergence time and 36%--49% in energy consumption compared to AsyncSGD.

We study privacy-preserving sparse linear regression in the high-dimensional regime, focusing on the LASSO estimator. We analyze two widely used mechanisms for differential privacy: output perturbation, which injects noise into the estimator, and objective perturbation, which adds a random linear term to the loss function. Using approximate message passing (AMP), we characterize the typical behavior of these estimators under random design and privacy noise. To quantify privacy, we adopt typical-case measures, including the on-average KL divergence, which admits a hypothesis-testing interpretation in terms of distinguishability between neighboring datasets. Our analysis reveals that sparsity plays a central role in shaping the privacy-accuracy trade-off: stronger regularization can improve privacy by stabilizing the estimator against single-point data changes. We further show that the two mechanisms exhibit qualitatively different behaviors. In particular, for objective perturbation, increasing the noise level can have non-monotonic effects, and excessive noise may destabilize the estimator, leading to increased sensitivity to data perturbations. Our results demonstrate that AMP provides a powerful framework for analyzing privacy-accuracy trade-offs in high-dimensional sparse models.

Open agentic systems combine LLM-based planning with external capabilities, persistent memory, and privileged execution. They are used in coding assistants, browser copilots, and enterprise automation. OpenClaw is a visible instance of this broader class. Without much attention yet, their security challenge is fundamentally different from that of traditional software that relies on predictable execution and well-defined control flow. In open agentic systems, everything is ''probabilistic'': plans are generated at runtime, key decisions may be shaped by untrusted natural-language inputs and tool outputs, execution unfolds in uncertain environments, and actions are taken under authority delegated by human users. The central challenge is therefore not merely robustness against individual attacks, but the governance of agentic behavior under persistent uncertainty. This paper systematizes the area through a software engineering lens. We introduce a six-dimensional analytical taxonomy and synthesize 50 papers spanning attacks, benchmarks, defenses, audits, and adjacent engineering foundations. From this synthesis, we derive a reference doctrine for secure-by-construction agent platforms, together with an evaluation scorecard for assessing platform security posture. Our review shows that the literature is relatively mature in attack characterization and benchmark construction, but remains weak in deployment controls, operational governance, persistent-memory integrity, and capability revocation. These gaps define a concrete engineering agenda for building agent ecosystems that are governable, auditable, and resilient under compromise.

Generalized Hopfield models with higher-order or exponential interaction terms are known to have substantially larger storage capacities than the classical quadratic model. On the other hand, associative memories for sparse patterns, such as the Willshaw and Amari models, already outperform the classical Hopfield model in the sparse regime. In this paper we combine these two mechanisms. We introduce higher-order versions of sparse associative memory models and study their storage capacities. For fixed interaction order $n$, we obtain storage capacities of polynomial order in the system size. When the interaction order is allowed to grow logarithmically with the number of neurons, this yields super-polynomial capacities. We also discuss an analogue in the Gripon--Berrou architecture which was formulated for non-sparse messages (see \cite{griponc}). Our results show that the capacity increase caused by higher-order interactions persists in the sparse setting, although the precise storage scale depends on the underlying architecture.

Autoregressive (AR) vision-language models (VLMs) have long dominated multimodal understanding, reasoning, and graphical user interface (GUI) grounding. Recently, discrete diffusion vision-language models (DVLMs) have shown strong performance in multimodal reasoning, offering bidirectional attention, parallel token generation, and iterative refinement. However, their potential for GUI grounding remains unexplored. In this work, we evaluate whether discrete DVLMs can serve as a viable alternative to AR models for GUI grounding. We adapt LLaDA-V for single-turn action and bounding-box prediction, framing the task as text generation from multimodal input. To better capture the hierarchical structure of bounding-box geometry, we propose a hybrid masking schedule that combines linear and deterministic masking, improving grounding accuracy by up to 6.1 points in Step Success Rate (SSR) over the GUI-adapted LLaDA-V trained with linear masking. Evaluations on four datasets spanning web, desktop, and mobile interfaces show that the adapted diffusion model with hybrid masking consistently outperforms the linear-masked variant and performs competitively with autoregressive counterparts despite limited pretraining. Systematic ablations reveal that increasing diffusion steps, generation length, and block length improves accuracy but also increases latency, with accuracy plateauing beyond a certain number of diffusion steps. Expanding the training data with diverse GUI domains further reduces latency by about 1.3 seconds and improves grounding accuracy by an average of 20 points across benchmarks. These results demonstrate that discrete DVLMs are a promising modeling framework for GUI grounding and represent an important step toward diffusion-based GUI agents.

Does Large Language Model (LLM) technology suggest a meta-semantic picture i.e. a picture of how words and complex expressions come to have the meaning that they do? One modest approach explores the assumptions that seem to be built into how LLMs capture the meanings of linguistic expressions as a way of considering their plausibility (Grindrod, 2026a, 2026b). It has previously been argued that LLMs, in employing a form of distributional semantics, adopt a form of holism about meaning (Grindrod, 2023; Grindrod et al., forthcoming). However, recent work in mechanistic interpretability presents a challenge to these arguments. Specifically, the discovery of a vast array of interpretable latent features within the high dimensional spaces used by LLMs potentially challenges the holistic interpretation. In this paper, I will present the original reasons for thinking that LLMs embody a form of holism (section 1), before introducing recent work on features generated through sparse auto-encoders, and explaining how the discovery of such features suggests an alternative decompositional picture of meaning (section 2). I will then respond to this challenge by considering in greater detail the nature of such features (section 3). Finally, I will return to the holistic picture defended by Grindrod et al. and argue that the picture still stands provided that the features are countable (section 4).

Technology change happens quickly such that new trends tend to crowd out the focus on what was new just yesterday. In this paper the peak popularity of the confluence of Object Technologies with early Web adoption is explored through the content of a seminar held in 1999. Distributed architectures were undergoing significant change at this point, and deeper software capabilities were just beginning to be broadly accessible over the Internet. The Object Web arose and was infused with new development tools reflecting these capabilities and allowing design of applications for deployment during the early days of the World Wide Web. This conference discussed the history, evolution, and use of these tools, architectures, and their future possibilities. The continued dominance of these approaches although under different names is demonstrated even though the term Object Web has receded in use. Favored newer offerings such as Kubernetes and microservices still model the core design attributes of the Object Web for example. Aside from connecting this seminar to relevance in the software world of today this paper also touches on the early AI tools demonstrated in this seminar a quarter century ago and how the popularity wave of any given technology might affect the current focus on AI technology offerings.

Creative approaches to attract students to software engineering at an early age are emerging, yet their differential impact on gender remains unclear. This study investigates whether design thinking's empathy-driven approach addresses the documented gender gap in interest in software engineering. In a 10-week curriculum-integrated design thinking software development course with 55 secondary school students aged 13-15 from two schools in Canada, we examined gendered differences in perceived gains in knowledge and interest, as well as in social-emotional experiences. Our results show that both girls and boys gained perceived knowledge in software development. However, girls showed significant improvements in self-efficacy, interest, engagement with sustainability topics, and well-being, including optimism, sense of usefulness, and social connectedness. Positive emotions were strongest during creative, collaborative phases, while technical tasks led to some boredom, especially among boys, though they still benefited overall. This suggests that human-centred design thinking might be one effective way to address gender equity challenges, though we need more differentiated technical implementations.

Interactive video generation has significant potential for scene simulation and video creation. However, existing methods often struggle with maintaining scene consistency during long video generation under dynamic camera control due to limited contextual information. To address this challenge, we propose MemCam, a memory-augmented interactive video generation approach that treats previously generated frames as external memory and leverages them as contextual conditioning to achieve controllable camera viewpoints with high scene consistency. To enable longer and more relevant context, we design a context compression module that encodes memory frames into compact representations and employs co-visibility-based selection to dynamically retrieve the most relevant historical frames, thereby reducing computational overhead while enriching contextual information. Experiments on interactive video generation tasks show that MemCam significantly outperforms existing baseline methods as well as open-source state-of-the-art approaches in terms of scene consistency, particularly in long video scenarios with large camera rotations.

Continual learning seeks to maintain stable adaptation under non-stationary environments, yet this problem becomes particularly challenging in object detection, where most existing methods implicitly assume relatively balanced visual conditions. In extreme-sparsity regimes, such as those observed in space-based resident space object (RSO) detection scenarios, foreground signals are overwhelmingly dominated by background observations. Under such conditions, we analytically demonstrate that background-driven gradients destabilize the feature backbone during sequential domain shifts, causing progressive representation drift. This exposes a structural limitation of continual learning approaches relying solely on output-level distillation, as they fail to preserve intermediate representation stability. To address this, we propose a dual-stage invariant continual learning framework via joint distillation, enforcing structural and semantic consistency on both backbone representations and detection predictions, respectively, thereby suppressing error propagation at its source while maintaining adaptability. Furthermore, to regulate gradient statistics under severe imbalance, we introduce a sparsity-aware data conditioning strategy combining patch-based sampling and distribution-aware augmentation. Experiments on a high-resolution space-based RSO detection dataset show consistent improvement over established continual object detection methods, achieving an absolute gain of +4.0 mAP under sequential domain shifts.