The Inference Report

Recent work on robot manipulation has focused on specialist policies trained on demonstrations. We show that generalist AI models can control manipulators more reliably than specialist policies such as vision-language-action models (VLAs) and world-action models (WAMs). We systematically analyze agentic manipulation policies (AMPs), in which a generalist AI model commands the manipulator directly, adapting its actions as the task unfolds. We present AMP1, a reference harness that runs unchanged across single-arm, bimanual, and mobile manipulation. AMP1 achieves state-of-the-art results on the RoboLab-120 and CaP-Gym benchmarks, with no demonstrations or task-specific training. On RoboLab-120, AMP1 reaches an 85.6\% success rate zero-shot, compared to 49.2\% and 42.9\% for leading VLA and WAM policies. The agent succeeds zero-shot in 104 of 108 real-world trials spanning twelve tasks and three robot instances, reducing failure rates by 20x and 15x relative to strong VLA and WAM baselines. The basic harness supplies no grasping, pushing, or placing skills; the agent develops its own strategies, such as searching for hidden objects, using objects as tools, and calibrating its cameras. With the addition of a persistent memory store, the agent can also learn from experience: an augmented harness with a memory representation that is carried across episodes lowers the failure rate on RoboDojo bimanual manipulation tasks by 4.6x relative to zero-shot deployment. Our analysis suggests that AMPs are substantially more general and robust than specialist manipulation policies, providing a promising foundation for robot manipulation.

Large language models reinvent the wheel all the time. They conflate task-solving with the execution of the solutions, which forces them to solve every problem from scratch. We show that once a task has been reasoned through, the application of its solution to a new problem instance can be done very efficiently and with little involvement from language models. We introduce agentic reasoning caching, a method for constructing robust CPU-executable agentic reasoning caches (ARCs) from LLM reasoning traces. ARCs show on-par performance with frontier LLMs across a range of difficult mathematical and scientific tasks, are fully explainable, continually trainable, robust to input perturbations, and 16-291x cheaper to invoke than the source model.

We prove that the set $\mathcal{N}_N$ of ReLU networks with one hidden layer and at most $N\in\mathbb N$ neurons is weakly closed in $W^{1,p}(Ω)$, for a bounded connected open set $Ω\subset\mathbb R^d$. Consequently, every coercive, weakly lower semicontinuous energy on this space attains its minimum over $\mathcal{N}_N$. This covers the Deep Ritz method with convex integrands, Sobolev training and Lipschitz learning. Moreover, as the weight decay parameter tends to zero, minimizers of the regularized problem converge, up to subsequences, to minimizers with minimal parameter norm. We also prove that the set of ReLU networks with two hidden layers of bounded widths is not closed in $W^{1,p}(Ω)$ or in $C(\overlineΩ)$. This disproves a conjecture of Petersen, Raslan and Voigtlaender.

Predicting the future motion of multiple people is inherently stochastic: the same observed history admits many plausible continuations depending on intent, coordination, and social context. Yet the dominant paradigm in multi-person 3D motion prediction remains deterministic, with existing methods producing a fixed number of outputs per agent. In contrast, we present a conditional flow matching framework that jointly generates diverse future motions for multiple agents, modeling both global trajectories and root-relative poses. A B-spline reparameterization compresses the generative state space while enforcing temporal smoothness, and a dual-stream transformer conditions each agent's predictions on observed neighbor motion for interaction-aware forecasting. Our method achieves state-of-the-art performance on CMU-Mocap (UMPM) and 3DPW, outperforming deterministic baselines on joint position error, pose error, and final displacement error despite producing a full predictive distribution. We additionally establish the first motion prediction baseline on WorldPose, a large-scale professional soccer dataset, where social conditioning demonstrably concentrates predictive uncertainty along interaction-constrained directions.

The Gaussian jet of an image is a graded vector space: a dilation by $λ$ multiplies derivatives of order $k$ by $λ^{-k}$. Local graded convolutions are exactly homogeneous differential operators, and a local feature is dilation-invariant if and only if it factors through the scale-normalized jet or, without the scale, through the retraction onto the weighted sphere of the grading. Under unknown Gaussian blur the retraction changes only by a change of scale, whereas the scale-normalized jet acquires a grading factor that no radial normalization removes. At the zero jet the retraction separates image families that radial normalizations and continuous classifiers cannot. Networks built on the retraction are universal for invariant features. On MNIST such a network keeps 88.4% accuracy under a two-pixel blur, against 66.2% for the same network on scale-normalized derivatives; in Lindeberg's Gaussian derivative network the retraction raises accuracy at that blur from 91.5% to 95.1%.

Test-time reinforcement learning (TTRL) adapts pretrained language models to unlabeled target-domain data using self-generated pseudo-labels, whose reliability remains poorly understood. We analyze majority-vote supervision for TTRL with online Direct Preference Optimization (DPO) and decompose its error into population-majority bias and prompt-dependent finite-rollout variation. Because population-majority bias is generally unidentifiable without external supervision, we focus on mitigating the downstream effects of finite-rollout variation. We propose \emph{Prompt-Conditional Distributionally Robust DPO (PC-Dr.DPO)}, which conditions within-prompt robust reweighting on observed vote reliability to make updates from sampling-sensitive prompts more conservative. We further introduce a cache-based mechanism to reduce self-supervised forgetting during online adaptation. Across six mathematical and knowledge-intensive reasoning benchmarks and five model backbones, PC-Dr.DPO matches or outperforms the strongest TTRL baseline in 24 of 30 settings.

Regression testing assesses whether software components that have already been developed and tested continue to work correctly after modifications. Ideally, regression testing re-executes the entire test suite after each modification, but this is infeasible in real-world scenarios because of the computational burden; therefore, optimized approaches, such as Test Case Selection, have been developed. The state of the art in regression testing optimization includes both classical approaches, such as Additional Greedy and DIV-GA, and quantum approaches, such as SelectQA, BootQA, and IGDec-QAOA. However, while the former incur high computational costs, the latter are constrained by the physical limitations of currently available quantum machines. This paper extends the evaluation of QAOA-TCS, our quantum approach for the test case selection problem. We thoroughly evaluate its performance, compare it with state-of-the-art classical and quantum approaches, and assess its robustness to quantum noise. The results show that QAOA-TCS outperforms state-of-the-art multi-objective approaches in most cases and maintains stable performance even under high noise levels. In the single-objective formulation, QAOA-based methods perform comparably. In terms of efficiency, our previous method based on quantum annealing, namely SelectQA, is the most efficient for large problems in the multi-objective formulation, whereas QAOA-TCS offers the best trade-off between effectiveness and efficiency in the single-objective formulation when run on realistic noisy simulators. However, such analysis is constrained by the different computational environments used to run the algorithms.

Large reasoning models (LRMs) improve task performance by generating longer intermediate reasoning traces, but this substantially increases inference costs and makes low-precision quantization increasingly important. However, aggressively quantized models often perform poorly on reasoning tasks. We observe larger activation fluctuations during the first decoding steps of reasoning. Calibration that omits these early states may miss their activation fluctuations. To address these reasoning-specific quantization challenges, we propose \textbf{CodeQuant2}, a clustering-based LUT quantization framework for low-precision large reasoning models. CodeQuant2 mitigates activation outliers with rotation and scaling, uses clustering to represent non-uniform weight distributions, and adapts codebook sizes to the quantization difficulty of each layer. Guided by our observation of reasoning-time activation behavior, CodeQuant2 further improves calibration sample selection to better match actual decoding statistics. Combined with our customized LUT CUDA kernel, CodeQuant2 delivers substantial accuracy improvements on reasoning benchmarks while achieving up to $2.85\times$ speedup over BF16.

Purpose: Surgical operative reports support procedural documentation, surgical training, performance assessment, and workflow review. However, extracting meaningful information often requires extensive manual review and clinical expertise. Automated surgical scene descriptions could reduce this burden by leveraging recurring surgical activities and prior knowledge. Methods: We propose GRAC (Graph-Based Retrieval-Augmented Captioning), a framework that represents surgical scenes as graphs capturing relationships among instruments, actions, and anatomical structures. Given an input scene graph, GRAC learns to retrieve relevant graph-caption pairs from training data. It compares instrument-tissue interactions between retrieved examples and the input scene, using comparison features to condition retrieved-caption representations during generation. Results: On the EndoVis-18 benchmark, GRAC achieved a CIDEr-D of 5.88, compared with 4.23 for SGT and 3.36 for the strongest conventional Transformer baseline. Exact instrument-action-target triplet-set accuracy reached 52.6%, compared with 32.7% for SGT and 26.2% for the strongest conventional baseline. These results demonstrate improvements in language quality and factual agreement. On TORS, GRAC achieved a CIDEr-D of 5.06, compared with 3.80 for the strongest baseline. Conclusion: GRAC demonstrates that combining structured surgical scene representations with retrieved examples improves surgical scene captioning and produces more contextually meaningful descriptions.

Loco-manipulation, which requires mobile manipulators to coordinate locomotion and manipulation over long horizons, is a key capability for large-object rearrangement in applications such as service robotics and warehouse automation. Despite recent progress, reliably executing these tasks remains challenging because performance depends on complex robot-object and object-environment interaction dynamics. Existing approaches address this challenge either through model-based methods, which often require accurate dynamics models, or through learning-based methods, which reduce this requirement but typically assume a fixed or predefined initial grasp. However, different geometrically stable grasps can lead to substantially different downstream manipulation performance. Therefore, we propose Grasp-TAGS, a task-aware grasp selection framework for goal-conditioned large-object rearrangement. We train a unified PPO-based reinforcement learning policy across multiple initial grasp configurations and distill the learned critic into a lightweight grasp utility network that ranks candidate grasps using only the candidate grasp configuration and goal information. At inference time, we select the grasp with the highest predicted downstream utility for the downstream policy to complete the task, before even making a contact. Experiments demonstrate that Grasp-TAGS improves rearrangement success over both fixed single-grasp policies and random grasp selection, achieving an average success rate of 89.84% compared with 79.87% for the best-performing single-grasp policies, an improvement of 9.97 percentage points. Grasp-TAGS also reduces final orientation errors across all evaluated objects. For more qualitative results, refer to the supplementary webpage: https://rpm-lab-umn.github.io/Grasp-TAGS-web-page/

Many multimodal alignment problems require mapping a source modality into a fixed space defined by a target modality, where the target space is not merely an intermediate embedding but is used directly for downstream analysis. In this setting, successful alignment should preserve both the global structure of the target distribution and the instance-level correspondence observed in paired source--target training data. Existing objectives typically address only part of this goal: contrastive losses encourage relative similarity, distribution-matching losses improve marginal agreement, and pointwise regression fits each source sample to its paired target without explicitly modeling target-space path geometry. We propose FlowAlign, a flow-guided target-space alignment method that repurposes conditional flow matching as a training-time alignment signal rather than an inference-time transport model. FlowAlign uses the learned flow field to guide the alignment of source-side representations with the fixed target space, encouraging them to preserve paired correspondence while remaining compatible with the global target-space structure. We further provide theoretical analysis showing how FlowAlign helps improve paired correspondence by controlling endpoint accuracy and capturing local target-space geometry. Experiments on synthetic data, paired single-cell multiome benchmarks, and image--text alignment demonstrate that FlowAlign improves instance level correspondence while preserving target-space structure across different modality settings.

Can a sender that only computes hashes, with a language model only at the receiver, transmit short human text more cheaply than a classical sender trained on the same domain? We answer with one pre-registered run on 1,000 English user utterances from Taskmaster-1. The primary endpoint is the cost per message to deliver 100% of the messages (hash bits, plus a 2-byte negative acknowledgement and the raw text when the first attempt fails), paired against an in-domain PPM order-1 sender. The hash-only sender delivered 857 of 1,000 messages on the first attempt and cost 192.8 bits per message, 18.4 bits more than PPM order 1 (95% bootstrap interval 10.9 to 26.3); by the pre-registered rule it spends more and is not equivalent within one byte. As secondary results we map measured sender memory against bits. The hash-only sender (8,190 B on a host build, keyed BLAKE2b tag included, as in every method) was the cheapest point in bits from 8,190 B up to, but excluding, PPM order 1 at 24,844 B, 45 to 46 bits below smaz2 and Unishox2; from PPM order 1 upwards, in-domain senders spend fewer bits, except order 2 pruned to 16 KiB. In airtime it loses to PPM order 1 on all five links priced, to Unishox2 on four, and to raw text with a tag on two, because each failure costs an acknowledgement frame and a fallback frame. A wider receiver search recovers half of the failures and makes the cost 7.7 bits lower than PPM order 1, at about three minutes of receiver time per failure, and still loses in airtime on four of five links. We report the measurement as a delimitation, with every deviation of the run.

In many areas of AI, general-purpose models have increasingly replaced handcrafted methods and specialized architectures. We argue that embodied AI is poised for a similar shift. Our key claim is that embodied tasks can be formulated as interaction with digital software. Specifically, we build software that renders an agent's view of a physical environment on a screen and provides controls for acting in it. A computer-use agent (CUA) can then perform embodied tasks through the same screenshots and keyboard and mouse actions that it uses to operate ordinary software. Using a single CUA design and changing only the software, our entries win first-, second-, third-, and sixth-place prizes across four ECCV workshop competitions, without any competition-specific agent tuning. These results suggest that the skills CUAs learn on software, such as visual grounding, precise clicking, and exploration, already carry over to embodied tasks, giving CUAs a far broader scope in embodied AI than previously realized. We therefore advocate that computer-use agents can be embodied generalists. Further, we advocate that, by following the training strategy that made computer-use agents general, embodied AI can converge on a single general model as well, ushering in a new era of general embodied AI. Website: https://pranjal2041.github.io/cua-embodied

This paper describes an approach and use case to estimate the correctness of LLM-generated COBOL summaries without access to human-authored references. The core intuition is that inter-model agreement on semantics is informative about truth, and that an LLM-as-judge can provide higher-fidelity pairwise assessments than traditional string- or embedding-based metrics. We operationalise this intuition with (i) controlled summary generation from diverse models; (ii) a LLM evaluator that compares pairs of summaries and emits structured judgments; and (iii) a calibrated aggregation rule that combines supportive evidence while penalising only corroborated contradictions.

The pigs were bleeding. That was the point of the experiment: could airway pressure reveal hemorrhage before it became critical? In the UCR Time Series Classification Archive, the de facto benchmark for Time Series Classification (TSC), the same recordings ask a different question: "which of 52 pigs produced this trace?". The dataset PigAirwayPressure is not an outlier. Although the archive is the standard benchmark, whether its tasks resemble the problems TSC is meant to solve has rarely been examined. We find that several datasets are repurposed from their original task, built from toy setups, or have flawed test splits. We further question how representative the archive's ordered non-temporal sequences (e.g. spectra, shape outlines) are of real TSC, and the potential downsides from the decomposition of multivariate problems. Our findings call for more careful interpretation of archive-wide scores and greater attention to deployment-relevant evaluation. The audit is at https://github.com/automl/ucr-provenance-audit.

Radiology AI has developed along two paths: specialized models that perform well on narrow tasks but cover few clinical findings, and generalist models that address a broader range of tasks but remain prone to hallucination and inaccurate measurements. Combining the two through agentic systems offers a route towards comprehensive radiology reporting, but published systems are tied to one modality or body part. We describe a radiology agentic framework that standardizes how an orchestrator calls modular tools to produce a structured report, and use it to measure what such a system can do with the tools from FDA-cleared radiology AI devices. Report generation is simulated on a large multi-modality report corpus, with the orchestrator given either a catalog of finding classes consolidated from the cleared devices, each available on the modalities it is cleared for, or an unlimited catalog in which any class can be requested, and planning either from the clinical indication or from the target report. Tool outputs are simulated from the reference report, so the comparison isolates tool availability and orchestration from tool error. With the cleared catalog and the indication alone, the orchestrator recovers under a third of the clinically actionable findings; recovery improves substantially with the unlimited catalog, and further when the plan is written from the target report. Attributing every missed finding to a cause shows that most are classes no cleared device offers for the study's modality, and that a third are classes or tools that were offered and not called. Adding the most frequently missing tools closes most of the gap to the unlimited catalog on held-out reports, whereas a smaller orchestrator with the unlimited catalog gives up the whole gain. Tool availability and effective orchestration are thus complementary constraints on comprehensive AI-assisted radiology reporting.

Reinforcement learning (RL) rewards are notoriously difficult to design and control, often leading to the model learning unintended behaviors such as reward hacking. One potential solution is to monitor for reward hacking and penalize it when detected; however, training against a monitor could lead to evasive behavior, and our general understanding of how to apply monitors effectively during training is limited. To study how best to use monitors to mitigate reward hacking, we introduce and open source three realistic environments where Qwen3-4B reward hacks: a coding environment hackable via test overwriting, a medical chat environment hackable via sycophancy, and a biography generation environment hackable via hallucination. We first focus on the coding environment, where we find that: (1) models can learn to evade highly accurate monitors by exploiting systemic flaws in probes and LLM judges; (2) monitors that leak more learning signal during RL suppress reward hacking but are more often evaded; and (3) including easier problems in training can decrease reward hacking. We apply our findings to build better reward hacking monitors for the medical chat and biography generation environments that improve upon naive baselines to reduce reward hacking rates across seeds to before-training levels in some interventions. Our results demonstrate that our takeaways translate to new settings and that better monitor intervention designs are possible.

The key-value (KV) cache dominates memory capacity and bandwidth in long-context LLM serving. While 4-bit microscaling formats (NVFP4, MXFP4) shrink the cache and run attention on FP4 Tensor Cores, a fundamental gap remains between algorithmic quantization and system-aware acceleration: existing FP4 KV systems use vanilla quantization that causes severe accuracy drops, whereas accuracy-oriented 4-bit algorithms rely on vector codebooks whose lookup overhead degrades decoding throughput. We present BitDecoding-FP4, an algorithm--system co-design for accurate and efficient native FP4 KV decoding. We mitigate three error sources: Dual-Residual Query Quantization splits queries into two FP4 tensors, cutting query logit error from 60% to 1%; Underflow-Aware Probability Scaling prevents E4M3 scales from rounding long-context probabilities to zero; and Outlier-Aware Key Quantization combines static smoothing, Hadamard rotation, and scale search to suppress key outliers under token-wise scaling. At the system level, we remove the softmax bottleneck via transposed softmax with fused probability quantization and stacked query MMA on paged KV caches. Across dense and MoE models with GQA and MLA, BitDecoding-FP4 stays within 0.5 points of BF16 KV accuracy with NVFP4 (1 point with MXFP4), and tracks BF16 KV in RL training rollouts. In end-to-end serving on NVIDIA B300 with Qwen3-4B and Qwen3-30B-A3B, it delivers 2.4-3.0$\times$ higher peak serving throughput than BF16 KV. The implementation is available at https://github.com/OpenBitSys/BitDecoding-FP4.

The timeline by which quantum computers can solve breakthrough problems in science depends on improved quantum algorithms as much as on better hardware. Such algorithmic advances are typically achieved through many years of expert human collaboration. In mathematics and computer science, large language models have accelerated progress through automated discovery loops. However, quantum algorithms can only be evaluated by classical computers on restricted or small instances. Here we demonstrate automated discovery of quantum programs beyond these regimes including order-of-magnitude improvements on scientific applications. To achieve this, we build a quantum program library in the Lean theorem prover in which large language model agents synthesize novel programs against formal problem specifications and verify their correctness and cost. Agent teams collaborate, sharing results and knowledge through wikis and discussing ideas at organized conferences. Our results improve on prior baselines for parallel search (up to 1.4x), quantum arithmetic (290x), and simulations relevant to fusion, photochemistry and drug-metabolism by 11x, 8x and 2.5x respectively. Combining human and artificial intelligence at scale, under the rigor of mathematical proof, lets researchers design algorithms that today's classical and quantum computers cannot test, shortening the path to practical quantum computation.

Dominant methods for interpreting large models impose strong structural assumptions. Generative Latent Priors (GLP) are a complementary approach: a diffusion model that learns the distribution of a frozen model's activations directly, imposing no assumptions. We port GLP to ESM-2 activations (UniRef50) and evaluate it at two host scales (ESM-2-8M and -650M), both as a lens for protein representations and as a prior for editing them. Generation quality scales smoothly with depth and compute. As a lens, single GLP "meta-neurons" are the strongest probes of discrete Swiss-Prot concepts at both hosts, beating InterPLM's released SAE features, and lead on continuous properties at 8M though not at 650M; falsification tests show this is mostly concentration, not new information. As an editing prior, the on-manifold correction makes DiffMean steering of three biochemical properties effective and restores on-manifoldness, but only in isolation: re-embedding the sequences actually produced after several rounds of decoding, the structural plausibility is compromised. This is mainly reproduced in additional SAE-based clamping structural steering experiments, where only Helix motif succeeds. The naturalness failure is traced to the rewriting coverage and the decoder rather than the prior, highlighting what the method must clear before it can serve for both discovery-by-editing and discovery-by-exploration. Trained GLPs are released on huggingface.co/collections/HPAI-BSC/glprot.

Machine learning (ML) methods have been explored as a means of optimizing lattice field theory calculations, with increasing interest in recent years. This chapter gives an introduction and summary to the use of ML methods to optimize lattice field theory without introducing any uncontrolled approximations. We focus on those directions which are clearly established as of late 2025, marking the 50th anniversary of the beginning of the study of lattice gauge theory. This chapter is a contribution to the online book ``Lattice QCD at 50 years,'' (LQCD@50), edited by Tanmoy Bhattacharya, Maarten Golterman, Rajan Gupta, Laurent Lellouch, and Steve Sharpe.

Uncertainty bands on wind forecasts help inform reserve sizing and commitments, but often fail to deliver their stated coverage. We apply conformal prediction to native bands to see what it adds and what it costs, comparing six forecasters, from ARIMA to a tabular foundation model, at two wind farms in Great Britain (GB) under operational conditions. We show that calibration repairs empirical marginal coverage with little or no Winkler Score penalty, but this apparent success conceals under-coverage in high winds. Calibrating by forecast wind regime turns marginal coverage into regime-level coverage, cutting day-ahead out-of-band volume in high winds by over 40% relative to pooled calibration while narrowing bands overall.

Autonomous coding agents are increasingly involved in software development, but their participation in real-world security remediation remains poorly understood. We investigate observable evidence of coding-agent involvement in commits referenced by reviewed GitHub Security Advisories (GHSAs), characterize the associated vulnerabilities, and examine review timing in relation to advisory disclosure routes. We develop a local-Git mining pipeline that identifies agent evidence through known agent accounts, explicit commit-message attribution, repository instruction files, and agent configuration artifacts. Linking these findings to GHSA-referenced commits yields 469 valid candidate advisories across 189 repositories, including 448 with direct attribution and 21 supported only by repository-level evidence. Claude is the most frequently observed tool family, followed by Copilot and Cursor. We further compare review timing against other advisories and historical baselines using an exact route--timing decomposition. Relative to advisories published in the same months, candidate advisories exhibit a modestly longer mean review lag, primarily associated with within-route timing rather than disclosure-route composition. Compared with a historical baseline, they exhibit substantially shorter mean review lags, with both route composition and within-route timing contributing to the difference. These findings demonstrate that coding-agent involvement is already observable across hundreds of security advisories and repositories, including high- and critical-severity vulnerabilities. By distinguishing direct attribution from repository-level adoption evidence and separating route-composition effects from within-route timing differences, we provide an empirical characterization of agent-associated remediation in the GHSA review pipeline.

A conversational agent that remembers has to decide what to keep. Retrieval-augmented generation keeps the transcript and searches it; the alternative is to compact each turn into atomic facts and search those. We compare the two on LoCoMo, with both arms answering all 1,985 questions in one window, against the same stores, with the same generator and judge, and both predictions on one row, which licenses McNemar's exact test on the 512 questions where they disagree. On the four categories comparable work reports, compaction scores 71.4 per cent against 64.6, a lead of 6.8 points (p = 5.5e-08) at 34 per cent fewer prompt tokens. On all five, including the adversarial set comparable work drops, the lead falls to 2.0 and stops being significant, because the baseline leads adversarial by 14.6. That loss has a structure. Our lead falls as the reference answer demands more source wording, 9.6 points ahead at one or two reference words and 3.1 behind past eight, with token overlap moving the same way, so the judge is not the cause. Compaction keeps what a fact asserts and discards how it was said, so it wins where the answer is a value and loses where it is a phrasing. We also report four negative results with their power calculations.

Graph similarity computation is an important task in graph mining with many real-world applications, such as chemical compound analysis, program dependence comparison, and social network analysis. The problem is computationally challenging: Graph Edit Distance (GED) and Maximum Common Subgraph (MCS), which quantify the distance between graphs and identify their largest shared subgraph, respectively, are NP-hard to compute exactly. To address these computational challenges, Graph Neural Networks (GNNs) have been applied to this problem. However, existing GNN-based approaches have primarily focused on small-scale graph datasets, while scalability to larger scales remains underexplored. In this work, we propose FiLSim, a scalable GNN-based framework, and investigate whether a simplified GNN pipeline can maintain competitive performance on standard benchmarks while scaling to large graphs. We additionally propose a sequential graph-editing strategy to construct large-scale graph datasets with reliable GED labels to evaluate scalability. The proposed framework is evaluated on four benchmark datasets and four newly created large-scale datasets, demonstrating its effectiveness and scalability. FiLSim remains accurate on graphs with more than 150{,}000 nodes, where the baseline methods run out of memory or fail to learn, and scores all test pairs within seconds on a single GPU.

LLM-based coding agents such as Claude Code are guided by repository instruction files (also called context files or agent configuration files) like AGENTS.md and CLAUDE.md, which are read in every agent session. As any documentation artifact, these files can fall behind the code and become stale (e.g., a test folder moves, but the file still points to the old one). In technical debt terms, such a stale instruction is documentation debt, and its interest is the cost it adds to every later agent session. Evaluations that check only task success can miss this cost, and the stale file can persist. In this study, we ask what leads agents to act on a stale instruction, how it affects their behavior and at what cost, and how they recover. We study three real-world stale instructions from active open-source repositories in depth, each a different kind of guidance: a test command, a code location, and a procedure for adding a component. Agents fix 22 maintenance tasks under three versions of each instruction (left stale, corrected to the current path, and removed), with the code and the task held fixed, and we analyze 732 trajectories of two models at the level of individual tool calls. We find that agents follow stale guidance when it is relevant to their tasks, sometimes even when correct information is available elsewhere. On the stale test command, Claude Sonnet 5.5 needed about two extra tool calls per run, a third more than with the corrected command; in the other cases, we could not detect a difference. Although agents generally recovered from the immediate failures, none repaired the underlying instruction file. Correcting the instructions eliminated operations on obsolete paths, while removing them prevented most. Our findings show that task success alone can hide the costs of documentation debt, highlighting the need to evaluate agent behavior beyond final outcomes.

Laya answers typed multiple-choice questions in one forward pass by scoring a hidden state at one [MASK] marker per option. We re-engineer the checkpoint (421M parameters; a 2-layer option head with no positional tensor; a documented but unused parallel layout) and test whether the answer depends on where an option is listed. Presenting 951 Qwen-verified questions under every cyclic rotation of their options (or six random permutations for 12 options), the released checkpoint changes its answer for 25.8% [22.9-28.7] of questions (19.7% on the 605 English-language items alone); accuracy with the gold option first versus last differs by 10.1 points, and the flips concentrate where the model's own confidence is low (54% versus 8% across confidence terciles). Fine-tuning with Laya's trainer on 2,818 teacher-verified rows with option shuffling cuts the flip rate to 7.2%; the identical rows without shuffling give 11.4%, so shuffling is responsible for 4.2 points of the improvement. A confidence-weighted vote over orderings gives 69.7% against 69.1% for a single base pass at 4.1x the compute. Labels are LLM-made and the test sets are teacher-filtered. We also fine-tune with the authors' parallel layout on the same rows: flip rate 0.5%.

Test-time scaling improves the performance of a fixed model by spending more computation at inference. There are two common ways to spend it on agents when a verifier can score every attempt. Best-of-N sampling runs many independent agents and keeps the best result but saturates quickly. One large team of collaborating agents can score higher but stops improving as it grows. Can we get the best of both? Our method, parallel agent teams, does so by sampling many small teams in parallel and keeping the best team's result. We test it on 12 long-horizon tasks from GPU kernels and formal proofs to exploit development. With a 27B model and a budget of 100 agents, parallel teams of ten beat parallel agents on 10 of 12 tasks and beat one team of 100 on all four tasks where we ran such large teams. At equal dollar cost the same parallel teams beat a frontier model on half the tasks. A simple statistical model shows that individual agents lose because their many samples have a low mean and that one large team loses because a single sample gets no gain from variance. Sampling many small teams in parallel balances a high mean with the gain from variance. Parallel agent teams thus offer a more effective way to scale test-time compute.

Current evaluations of agentic AI systems primarily rely on benchmarks. However, the static nature of benchmarks makes them uninformative about the value of auditing interventions at deployment time. For reliable deployment, we often need to determine within an agent's trajectory whether or not to intervene, e.g., by verifying the agent's current output or steering it toward different behavior, using either human review or more costly models or time-consuming agentic supervision. The value of such interventions is fundamentally a causal inference question, asking whether the causal effect of such an intervention is beneficial for the final outcome. Estimating this causal effect is particularly challenging due to the dynamic nature of the problem: the impact of an intervention depends on the current context and environment, as well as the agent's future behavior. We develop the first methodologically grounded framework for evaluating the causal effects of randomized interventions within agent trajectories. By leveraging a relatively small dataset of agent trajectories collected under randomized interventions such as verification or screening, we can estimate the dynamic causal effect of a rich class of (contextual) intervention policies. The structural parameters that we estimate can also be used to construct safety leaderboards that capture the risk of various agentic actions, which can be used by practitioners for audit prioritization. Applying our methods to tau-bench Retail, ToolSandbox, and ImpossibleBench, we find that intervention effects vary substantially across agent actions and states. More broadly, our work suggests a rich research agenda at the intersection of causal inference and agentic evaluation, which we believe is essential for the reliable and safe deployment of AI agents in the real world.

Competitive programming problems pose distinctive challenges for Large Language Models, as they demand not only proficient code generation but also advanced algorithmic reasoning. This study presents a Retrieval-Augmented Generation approach specifically designed to enhance the capacity of Large Language Models to address such problems. Our approach integrates an embedding-based retrieval mechanism and a similarity search algorithm to identify analogous problems within a knowledge database, employing a dynamically adjusted threshold to mitigate retrieval noise. We conducted an evaluation using the CodeContests dataset, focusing on the test split problems that exhibit similarity to instances from the training split. The experimental framework considered three conditions: (i) problems provided solely with their original descriptions, (ii) problems supplemented with randomly selected examples from the training split, and (iii) problems augmented with similar examples retrieved through the proposed method. In addition, we incorporated a Min-K membership inference analysis to examine whether the generated code solutions were influenced by potential memorization of training data within the evaluated Large Language Models. This complementary perspective enables a deeper understanding of model behavior by distinguishing genuine generalization through retrieval from implicit recall of pretraining samples. Performance was assessed using both pass rates and the CodeForces Elo rating of the problems. The results provide empirical evidence that similarity-based retrieval substantially enhances Large Language Models performance in competitive programming scenarios, offering an scalable framework for improving code generation in algorithmic problem-solving contexts.

External lexical resources can help large language models (LLMs) translate texts in low-resource languages and specialized domains, but these resources may conflict with models' prior lexical knowledge. To isolate this conflict, we construct a controlled Mandarin-to-English translation task in which familiar source words are assigned new meanings within and across semantic categories. We treat these prompt-supplied meanings as contextual knowledge (CK) and systematically evaluate the effects of word frequency, sentence context, and prompt framing on the translated outputs. Across nine models, CK reliance is greater for less frequent source words and in sentence contexts that support the supplied meanings. Prompt framing can reduce CK reliance, but seven of nine models favor CK even when explicitly instructed to ignore the supplied meanings. Within model families, larger models are more likely than smaller models to utilize CK in their outputs. Three follow-up experiments further show that modal weakening lowers CK reliance, the effect of repetition depends on prompt layout, and the effect of reasoning varies across models. Overall, our controlled study reveals how CK reliance varies across conditions and models, motivating new methods to detect and evaluate lexical conflicts and guide models' lexical choices in LLM-based translation.

Leaderboards are a popular way to compare large language model (LLM) capabilities and can influence how model developers invest resources. We ask when competition on leaderboards leads developers to produce models with similar capability profiles, which we call monoculture, or similar performance levels. We study this by formulating leaderboard competition as a game in which model developers invest resources across capabilities to maximize their win rate. We show that similar production functions can induce similar capability profiles, while comparative advantage can sustain distinct capability profiles. We then study convergence in performance levels and show that diminishing returns can lead to an increasingly crowded competitive frontier. Using data from LM Arena and OpenRouter, we find that developers improve more in categories where they lag competitors, while still retaining category-specific advantages. At the same time, we find that the frontier of models becomes increasingly crowded, alongside increasingly diversified model usage patterns. Our results suggest that increasing competition does not need to produce interchangeable models, but can instead sustain competitive models with distinct strengths.

We study planning for finite-horizon partially observable Markov decision processes (POMDPs) with finite state, action, and observation spaces. Both exact and approximate planning are known to be computationally hard in this setting, so prior work has explored natural assumptions that make planning easier. Recent work [Golowich et al., 2022] gave quasipolynomial-time approximate planning algorithms for observable POMDPs. These are POMDPs whose observation channels satisfy an $\ell_1$-observability condition, but notably, no assumptions are required on the transition dynamics. In a similar vein, for every fixed $ρ>0$, we give quasipolynomial-time approximate planning algorithms for $ρ$-decodable POMDPs. We show that decodability is implied by both $\ell_1$- and $\ell_2$-observability and is strictly weaker than observability. Moreover, we show that the optimal decodability constant can be computed in polynomial time, whereas computing the optimal $\ell_1$-observability constant is NP-hard in the worst case. We also show that the fast filter-forgetting guarantee underlying the approach of [Golowich et al., 2022] does not extend directly from $\ell_1$-observability to constant $\ell_2$-observability. Thus, decodability is an easy-to-check, unifying assumption under which quasipolynomial-time approximate planning for POMDPs is possible.

Multimodal Large Language Models (MLLMs) require continuous factual updates to stay current, yet lifelong knowledge editing remains a daunting challenge. The primary obstacle is the severe Locality Erosion and Catastrophic Forgetting, where sequential edits inevitably interfere with pre-trained capabilities and prior updates. In this paper, we propose \textbf{ROSE}, a \textbf{R}isk-aware \textbf{O}rthogonal \textbf{S}ubspace navigation framework for lifelong knowledge \textbf{E}diting. ROSE addresses forgetting by aiming to restrict parameter updates to the orthogonal complement of previously edited subspaces. Crucially, it introduces a risk-guided mechanism that utilizes a dynamically computed risk mask to safeguard specific factual parameter subspaces essential for model stability. During inference, ROSE employs a multi-granularity knowledge integration strategy, featuring prototype-anchored semantic gating and perturbation-driven subspace fusion to precisely activate edited knowledge for relevant queries while limiting interference on unrelated inputs. Comprehensive experiments on two prominent multimodal benchmarks show that ROSE is competitive with state-of-the-art approaches across five key metrics, while maintaining outstanding stability over $1,000$ sequential updates. Our code is available at https://github.com/sunlitsong/ROSE.

Animation is pervasive on the Web, the CSS animation property alone appears on over three-quarters of webpages. Yet existing webpage code generation benchmarks target static screenshots, user-triggered interactions, or isolated motion graphics, leaving the ability of multimodal large language models (MLLMs) and coding agents to reconstruct complete animated webpages largely unexplored. To fill this gap, we introduce AnimWeb-Bench, the first benchmark for animated webpage code generation from videos. It comprises 124 real-world webpages spanning 7 application domains and 33 animation types, with 91.9% of pages combining multiple types, each paired with a reference video and the static assets of the original webpage. We further propose a human-aligned, two-dimensional evaluation protocol: Static Appearance Fidelity measures structure, text, color, and visual similarity on aligned keyframes, while Animation Fidelity tracks UI components across frames and compares their position, size, velocity, and appearance with those in the reference video. Evaluating 14 MLLMs and three coding agents reveals: (1) reproducing webpage animations is substantially more difficult than reproducing static appearance; (2) explicit frame inspection and render feedback improve the reconstruction quality of coding agents; and (3) rapid, continuous changes are harder to reproduce than discrete state transitions. Our code and dataset are available at https://starry1001.github.io/AnimWeb-Bench-website/ .

Large language model (LLM) agents coordinate through conversation, and the same message can be phrased in simpler or more complex language, bluntly or politely. We test whether such differences change how other agents judge and act, using the LURK Werewolf framework, which separates each utterance's speech-act predicate from its natural-language realization. We rewrite only the realization, check that the game content is preserved, and replay each game day from saved checkpoints against a re-sampled control. We vary linguistic range consistently across a speaker's messages and politeness strategy per utterance. Across 30 games and 4,335 replays, neither manipulation changes the trust other agents place in the speaker, and no effect on votes or night actions survives correction for multiple comparisons; the largest uncorrected effect is that shifting range away from a neutral paraphrase changes about 4 percentage points more of the other players' votes. In natural play, 88% of face-threatening speech is bald on-record regardless of role.

Probably Approximately Correct (PAC) bounds on the generalization error (the difference between the empirical prediction error computed on the training data and its expectation) are an established tool in statistical machine learning, and there have been several extensions of these bounds to dynamical systems. In parallel, in recent years, there has been growing interest in finite-sample bounds for parameter estimation error in system identification. The latter bounds tend to be algorithm-specific, relying on reformulating the identification algorithm as a least-squares regression problem. This is in contrast to classical PAC bounds, which are algorithm-agnostic and can be applied to any identification algorithm based on minimizing a cost function involving the empirical loss. In this paper, we show how oracle bounds derived from PAC bounds can be used to derive finite-sample bounds for dynamical systems. As an illustration of this approach, we present finite-sample parameter-estimation bounds for linear and bilinear systems with bounded inputs and noise. These bounds apply to identification algorithms based on minimizing a regularized empirical loss function. The key ingredient is to establish an oracle inequality, i.e., an inequality between the true prediction error of the learned model and the minimum achievable true prediction error, and then use either the convexity of the true loss or system-theoretic properties of the loss function to derive bounds on the parameter estimation error. We also analyze the rates of the obtained bounds and compare them with those of classical asymptotic results.

Chain-of-thought (CoT) reasoning has emerged as a powerful paradigm for extending the computational capabilities of transformer-based models. A growing body of recent work has begun to uncover the fine-grained computational power of transformers augmented with CoT, as well as their connections to classical complexity theory. In this work, we advance this line of research by establishing the first hierarchy theorem for CoT-based computation. Our main result shows that for any reasonable resource bound $r(n) \geq n$ and any $\varepsilon > 0$, \[ \mathrm{CoT}(r(n)) \subsetneq \mathrm{CoT}(r(n)^{1+\varepsilon}), \] where $\mathrm{CoT}(r(n))$ denotes the class of decision problems solvable by transformers using $O(r(n))$ CoT reasoning steps on inputs of length $n$. Our result shows that increasing the number of CoT reasoning steps provably yields strictly greater computational power, establishing CoT steps as a fundamental resource in transformer-based computation. Unlike classical hierarchy theorems, which are typically proved via diagonalization, our approach leverages known coarse-grained simulations between CoT-augmented transformers and Turing machines in both directions. Building on these connections, we adapt the classical padding technique from complexity theory to this setting to obtain a tight hierarchy. A key technical contribution of our work is the implementation of padding within the transformer-based CoT model.

The success of modern score-based generative models is largely driven by diffusion-based training, which learns a collection of time-indexed score fields corresponding to progressively noisier versions of the data distribution. While such models achieve impressive sample quality, leveraging them for sound probabilistic inference---particularly sampling from arbitrary conditionals of the model---remains challenging. This difficulty arises because conditioning information is only available for clean data and not for the higher noise levels required to generate exact conditional samples. In this paper, we introduce DIffusion-free SCOre matching (DISCO), a principled approach that sidesteps the need for time-dependent score fields altogether. Our method is based on an objective that, while reminiscent of diffusion-based training, estimates only a single score of the (slightly perturbed) data distribution, for which the exact conditional score is readily obtained. In experiments, DISCO is competitive with state-of-the-art diffusion models in terms of sample quality. More importantly, DISCO yields a more faithful representation of the underlying data distribution and---crucially---enables accurate sampling from arbitrary conditionals, outperforming standard heuristic samplers. This capability opens the door to principled and flexible probabilistic reasoning with score-based models.

Extracting discriminative features from laser-induced breakdown spectroscopy (LIBS) spectra is important for the classification of complex biological samples. This study evaluates continuous wavelet transform (CWT)-based spectral peak extraction for distinguishing human plasma samples labeled as SARS-CoV-2-positive or negative. Plasma specimens from 181 individuals, comprising 53 positive samples and 128 negative controls, were deposited as 5 μL droplets on silicon substrates and dried before analysis. Spectra were acquired using a nanosecond Nd:YAG laser operating at 1064 nm, with 100 single-shot measurements collected per specimen and averaged after intensity-based outlier rejection. CWT-based peak detection was coupled with a support vector machine (SVM) classifier. The reported leave-one-out evaluation yielded an accuracy of 93.00%, sensitivity of 94.67%, and specificity of 91.61%; these estimates remain provisional pending reconciliation with specimen-level predictions. Features identified by the CWT pipeline included emission lines assigned to Ca II, K I, and N I. The workflow provides a basis for examining spectral features associated with the studied plasma groups. Independent validation and assessment of specimen, substrate, and acquisition-related variability are required to establish generalizability and diagnostic relevance.

Layered architecture recovery (LAR) reconstructs a system's architectural layers from its package dependencies, and is defined as a combinatorial optimisation problem solved with classical search. We investigate whether the Quantum Approximate Optimization Algorithm (QAOA) can be applied to LAR, and with what practical limitations. We formulate LAR as a Quadratic Semi-Assignment Problem, transform it into a Quadratic Unconstrained Binary Optimisation (QUBO) problem in which each package is penalised unless it is assigned to exactly one layer, and encode it as an Ising Hamiltonian solved by QAOA. To reach instances beyond the qubit budget of a single circuit, we also apply an impact-guided decomposition that solves the problem as a sequence of smaller sub-problems. We evaluate the approach on three real Java systems ($15$, $18$, and $24$ qubits), studying the effect of circuit depth and sampling budget (RQ1), a comparison with a classical Tabu Search baseline (RQ2), robustness to simulated quantum noise (RQ3), and the decomposition (RQ4). We find that monolithic QAOA is feasible and accurate on the two smaller systems but fails on the $24$-qubit instance, that the sampling budget rather than circuit depth governs its accuracy, and that it is competitive with but does not outperform Tabu Search. Decomposition restores reliable recovery on the $24$-qubit instance, raising the optimal rate from $6.7\%$ to $93.3\%$; its accuracy is governed by the size of each sub-problem and, unlike the monolithic pipeline, improves as the sampling budget is reduced. Our results indicate that, in the current noisy intermediate-scale regime, problem decomposition, rather than larger circuits or more sampling, is the key to applying QAOA to software architecture recovery.

AI agents can execute multi-step tasks on users' behalf, yet the boundaries of acceptable delegation remain unclear. We report a scenario-based survey of 289 U.S. adults with prior AI experience covering 45 use cases across nine life domains. Each respondent evaluated one scenario per domain, producing repeated measures of automation preference, trust, collaboration, AI leadership, privacy concern, and willingness to pay. Automation preferences varied substantially across domains. Generalized estimating equation models showed that greater trust in the AI agent, desired AI leadership, prior task-specific AI use, and anticipated financial savings were associated with greater automation preference. In contrast, concerns that automation would reduce personal fulfillment predicted lower automation preference. Open-ended responses further revealed a preference for bounded autonomy, with participants often favoring substantial AI execution while retaining human involvement, oversight, or final authority. We argue that delegation is best understood as a configuration of execution, collaboration, leadership, and authority.

Large language models (LLMs) have shown promise for automating data science workflows. Benchmarks such as DataSciBench (2026) and StatQA (2024) have advanced the evaluation of their capabilities across data science and statistical tasks. Assessing their reliability requires examining both the quality of their responses and their consistency across repeated runs, task types, and difficulty levels. With this aim, we introduce DS-QA, a benchmark comprising 100 data science and statistics questions drawn from UK university courses, spanning 16 datasets and three difficulty levels. We develop a three-layer evaluation framework that combines general response properties (runtime, verbosity and text similarity), coarse-grained metrics (accuracy, code executability and completeness), and task-specific metrics. We evaluate OpenAI models using a Monte Carlo protocol with 100 repeated runs. The results reveal substantial variability in responses to data visualisation and supervised learning tasks, particularly at medium and hard difficulty levels, highlighting the importance of evaluating consistency alongside response quality. We also find that an increase of 0.1 in the word verbosity ratio corresponds to an increase of about 0.013 in the token verbosity ratio.

We propose a fixed-point acceleration method for optimization over probability measures, applicable to a variety of machine learning problems. We represent probability measures as pushforwards of a reference distribution through a neural network map and equip the parameter space with the induced pullback Wasserstein metric. This geometric structure allows natural gradient descent to be interpreted as a fixed-point iteration on a Riemannian manifold. We accelerate this iteration using Riemannian Anderson mixing method, which exploits historical residuals to capture approximate second-order information without forming Hessians. In the present parametric setting, the manifold is represented by Euclidean coordinates endowed with a point-dependent metric, simplifying retraction and vector transport. We evaluate the method on multimodal sampling, PDE-based inverse problems, aggregation dynamics, Bayesian logistic regression, and generative modeling. Across the tested tasks, Anderson mixing reduces the number of natural-gradient iterations and achieves competitive performance relative to SGD and Adam in terms of both loss-gradient evaluations and wall-clock time. These results suggest that fixed-point acceleration is an efficient tool for improving the computational efficiency of training on parametric Wasserstein methods.

Large-language-model decode attention streams the key-value (KV) cache for every generated token. Packing that cache into 2-bit or 4-bit quantized integers reduces memory traffic but does not automatically translate into latency saving. Existing 2-bit fused kernels reach under 31%/15% of peak HBM throughput on H100/B200. Converting each value and applying its scale and offset before feeding Tensor Cores is a major bottleneck because instruction-issue throughput is scarce on modern GPUs. Instead of applying scale and offset to each KV value, FOSCAR uses affine folding to apply them to the other operand or the output of decode-attention matrix products, depending on the quantization scheme (token-wise or channel-wise). This reduces scale-and-offset operations by up to two orders of magnitude while preserving the dequantized attention output in exact arithmetic. FOSCAR also uses bit manipulation to feed KV operands to Tensor Cores efficiently, requiring only 1 and 0.19 ALU instructions per INT2 element on H100 and B200, respectively. Finally, FOSCAR builds warp-specialized GPU pipelines for Hopper (H100) and Blackwell (B200) using wgmma and tcgen05. On H100, FOSCAR INT2 and INT4 kernels achieve geo-mean speedups of $3.36\times$ and $2.83\times$ over FlashInfer BF16, and $2.03\times$ and $1.71\times$ over TensorRT-LLM FP8. On B200, the corresponding INT2 and INT4 speedups are $2.67\times$ and $2.30\times$ over FlashInfer BF16, $1.48\times$ and $1.28\times$ over TensorRT-LLM FP8, and $1.16\times$ and $1.00\times$ over TensorRT-LLM NVFP4. We integrate FOSCAR into SGLang to benchmark serving efficiency. On long-context serving workloads, FOSCAR increases peak aggregate throughput by up to 170% and 341% over 2-bit and 16-bit baselines, respectively. We open-source FOSCAR as a kernel library at https://github.com/FutureMLS-Lab/FOSCAR.

LLM agents benefit from using memory systems to improve performance on long-horizon, multi-session tasks. However, memory introduces a critical attack surface. Malicious instructions encountered in one session may be stored and propagated through memory before being retrieved and acted upon in later sessions. Because memory systems differ in how they process and retrieve information, the downstream effects of poisoning can vary significantly across systems. To quantify the effects of poisoning across different memory architectures, we introduce MemoryDojo, a benchmark for tracing the lifecycle of poisoned information through agent memory and measuring its influence on subsequent agent behavior. MemoryDojo is easily extensible, consists of 155 multi-session long-horizon trajectories covering four domains, and supports five different memory systems. Through our evaluation, we find that all memory systems are susceptible to memory poisoning, with substantial variation in downstream effects. We further propose a memory-aware poison optimization process that increases the poisoning efficacy, highlighting differences between memory architectures. We hope that MemoryDojo will support research on understanding and mitigating agentic memory risks.

Modern LLM serving systems use prefix caching to accelerate multi-turn conversational workloads. They cache the key-value states of previous conversation prefixes and reuse them when subsequent requests share the same prefix, to avoid redundant prefill computation and reduce serving latency. However, achieving a high cache hit rate remains challenging because cache reuse depends on whether a conversation continues and how soon the next turn arrives. Our analysis of real workloads shows that user intent is a useful signal for this behavior, as different intents exhibit different continuation probabilities and inter-turn delays. Based on this key insight, we propose Intent-Aware Caching (IAC), an online eviction policy that classifies requests by intent, learns lightweight reuse statistics from recent serving logs and leverages these statistics to guide an oracle-inspired eviction strategy. Extensive evaluation on real workloads (WildChat and ShareChat) shows that IAC reaches a 34.6% cache hit rate. This substantially outperforms both the previous learning-based approach (18.8%) and traditional methods such as least recently used (LRU, 12.2%), while reducing average serving latency by up to 46.1%.

Digital twins of intensive care patients should update forecasts from new data and predict treatment responses, but few studies test whether routine records support either. We propose ICare-DT, a Transformer-based twin that updates a latent patient state hourly from new measurements without retraining. We trained it on 140,473 eICU-CRD stays to forecast vital signs crossing their training-set means. On test data, later forecasts beat the twin's own hour-6 forecasts of the same windows in all 12 settings (+0.073 to +0.153 AUROC). On validation data, controls that withheld or swapped the new hours traced the gain to the patient's own measurements. Without refitting, the model kept its gain on a second database (8 of 9 settings, +0.025 to +0.114 AUROC) in a prespecified evaluation of the previously analyzed MIMIC-IV test split. Refitted within MIMIC-IV, a twin state fine-tuned on clinical endpoints gave gradient boosting a significant gain for acute kidney injury (+0.0035 AUROC; hypotension +0.0010, not significant). To test response prediction, we built a benchmark that replaces the treatments in real records with simulated doses and injects responses of known size, visible only at measurement times. Ridge models supervised with injected effects recovered them from real records only through frequently measured variables (heart rate) read at their measurement times. We asked standard learners to predict these effects for a dosing regime no patient had received. No standard learner outperformed predicting zero effect at an out-of-support query. When told how the simulated responses were generated, a separate learner recovered the effects from the same records without seeing their true values (R2 0.30 to 0.40 at one effect size). Routine records support forecasts that improve with each patient's new data, and our benchmark is a testbed for models that learn how responses are generated.

RidgeFill shortens prefill by not computing most of a model's upper layers for the prompt. The lower layers run exactly up to a split; from the split layer's state a per-layer map predicts every later layer's input, and each late layer writes its cache from that prediction. The design follows YOCO and DeepSeek-V4.1-Flash, where the projection is trained into the model. RidgeFill retrofits it onto models that were never trained for it: the map is a ridge regression fitted once in closed form from one pass over calibration text, the layer's own projections run unchanged, and a chosen set of rows, the last few thousand tokens, and every decoding step run the full model. Nothing is trained and no token is dropped. Most of the work was in measurement, finding a repeatable methodology. Loss averaged over all tokens hides RidgeFill's error. Needle retrieval in the filled region, loss on tokens that depend on distant context, and splice tests that swap filled keys or values for exact ones show which component carries it: the keys on Llama-3.1-8B, half of it in their scale, and the values and the recurrent state on Qwen3.5-9B. On Qwen3.8-27B in a llama.cpp fork on an A100, time to first token is 1.25 to 1.31x faster at 16K to 120K tokens with the maps in VRAM; RULER changes by -1.2 and -1.4 points and LongBench v2 by -1.4; the 95 % intervals include zero but do not rule out losses of a few points. On Qwen3.8-Flash-Next in vLLM it is 1.50 to 1.60x faster than exact at 9K to 32K tokens, with RULER's needle tasks at 99.8 against 100.0 and perplexity on novel tokens 2.65 % higher [-0.02 %, +4.93 %]. On three 8 to 9B reference models, at the same split and tail as the identity fill, the fitted map retrieves 49.1 and 49.6 more filled-region RULER points on the two dense models (at that split Llama's loss is worse), and its long-document loss is +/-0.03 nats of the released SwiftKV checkpoint's.

Anticipating Q-day security concerns, major CDNs and cloud providers are migrating to post-quantum cryptography (PQC). While existing literature benchmarks PQC performance, empirical research evaluating real-world adoption data remains scarce. We present an empirical assessment of PQC deployment across seven critical protocols (HTTPS, LDAPS, SMTPS, IMAP-SSL, POP3-SSL, TELNETS, and RDP) using 10,000 target IP addresses per protocol. At the origin-server level, PQC algorithm support reaches 9.1% for HTTPS and 7.39% for LDAPS, while falling below 1% for all other evaluated protocols. At the entity level, PQC support is 11% for HTTPS and 5.4% for LDAPS, while the remaining protocols have less than 5% of entities supporting PQC on the hosted servers. We further correlate these adoption trends against organizational digital and financial footprint, and industry sectors, yielding a multi-dimensional empirical assessment of PQC deployment. Finally, at the domain level, we perform a comparative analysis of PQC adoption in Cloudflare-hosted domains against baseline Cloudflare datasets. Collectively, our findings provide critical insights into the current state of PQC readiness over Transport Layer Security (TLS).

With the anticipated arrival of Q-day, quantum computers threaten current cryptographic schemes for encryption and digital signatures. In response, major content delivery networks and cloud service providers are already adopting Post-Quantum Cryptography (PQC). The current state-of-the-art focuses on benchmarking the operational performance of implementing PQC, but the real-world adoption status remains under-researched. This work utilizes large-scale SSH fingerprint data from a proprietary scanning tool to analyze PQC adoption across 1.52 million hosts and 27,012 entities. Our study shows that the PQC transition in the SSH protocol is underway but highly polarized: 45.2% of SSH hosts on port 22 support PQC, while 54.8% remain classical-only. At the entity level, 39.9% have enabled PQC on at least one host, whereas 60.1% have no PQC support. By correlating PQC adoption with industry, geography, and organizational size (digital and financial), this study provides a comprehensive empirical assessment of PQC adoption status in SSH infrastructure.

Language-model systems can select prompts and parameter adapters repeatedly during deployment, using feedback to guide later choices. Earlier choices may change the context or environment inherited by later configurations. We study selection among preconstructed candidates and compare cumulative loss with the best fixed prompt--parameter pair used from the beginning. This excess loss separates into three terms: selection regret within the available family, the gap from restricting that family, and the benefit or harm of the actual deployment history. Selection is evaluated using the history each candidate would have produced if used throughout, which separates candidate quality from the effects of switching. When loss functions are fixed independently of selector randomness, past choices affect loss for only a bounded number of rounds, and every candidate's own-history loss is revealed, limiting switches yields sublinear expected regret within the selected family. We also quantify the costs and additional assumptions of independent probing and deployment-only feedback. Controlled experiments show that fewer switches can either lower or raise deployed loss, depending on whether inherited histories are harmful or helpful. With Qwen3.5-9B on Spider, text-only selection has lower within-family regret than joint selection, yet produces more failures because its candidate family excludes stronger configurations.

Attention is a major scalability bottleneck in large language models (LLMs), with its computational cost and key-value (KV) memory traffic becoming increasingly prohibitive as context length grows. Sparse attention addresses this bottleneck by restricting each query to a subset of the context, but its effectiveness hinges on identifying relevant regions both cheaply and reliably. Existing training-free methods primarily rely on runtime relevance estimates, with selection dimensionality and granularity typically governed by fixed heuristics rather than the intrinsic interaction structure of pretrained attention heads. We take a different perspective: the query-key operator of each pretrained head defines a characteristic response structure that determines the resolution needed for sparse selection, both across feature directions and across context tokens. Building on this insight, we propose Head-Adaptive Sparse Attention (HASA), a training-free sparse attention framework that jointly adapts selection resolution along these two dimensions. Along the feature dimension, the singular spectrum of each head's query-key operator determines an effective rank, yielding a compact, head-specific response space. Along the sequence dimension, neighboring regions are progressively merged according to the directional coherence of their token responses in this same space, while heterogeneous regions remain separate. Exact attention is then computed using the original full-dimensional KV states of the selected regions. HASA captures 88.1% of dense-attention mass and, compared with UNIQUE, a strong recent sparse-attention baseline, improves accuracy by up to 3.47 and 2.57 points on RULER and LongBench-Pro, respectively, and accelerates self-attention by up to 1.32x, while achieving up to 14.07x speedup over dense attention.

LLM-guided search loops propose a modified candidate, score it, and keep it if it scores better. When both candidates are re-scored at every comparison, as in the minibatch acceptance gates of prompt optimizers, keep-if-better is a reversible Markov chain if the score difference is logistic and the mutation operator is reversible, and its stationary law is the mutation equilibrium tilted by exp(beta f). For any continuous symmetric noise difference the same tilt is the stationary law of a small-step diffusion limit, with beta = 4F'(0); for a minibatch gate beta grows with the square root of the minibatch size. Caching the incumbent's score instead inflates the recorded score. We test the tilt as a prediction for LLM-guided program search, committing each prediction from calibration data before the search runs. The level after 150 steps rises with the minibatch size in six of seven settings, but its predicted value fails: on bin packing, search attains 62-88% of the predicted gain when the proposer edits the parent and 109-119% when it writes new heuristics. The attained fraction falls with the parent-child quality correlation of unselected mutation; a registered test on two new settings confirmed the direction, but not the quantitative agreement after a correction. Further registered tests found that runs started from a good program kept losing quality over 300 steps, more so from starts that search had selected, by a margin that passed its registered rule narrowly. Runs from different starts had not reached a common level, so whether the tilt describes the stationary law of LLM editing remains open. A registered test in prompt optimisation on frozen task-model scores failed: search exceeded the tilt, and the bin-packing relation did not transfer.

Vision-language-action (VLA) policies have become a promising paradigm for end-to-end robot control, but their effectiveness still depends heavily on large volumes of demonstration data. In manipulation tasks, this dependence is compounded by structured spatiotemporal uncertainty in the action space: multiple nearby actions or trajectories may successfully solve the same task, while semantically equivalent demonstrations often unfold at different speeds. Standard pointwise supervision treats each demonstrated action as an exact target, imposing an unnecessary learning burden under such uncertainty. We present SURE, a data-efficient VLA training framework that explicitly models both spatial and temporal uncertainty. SURE augments a VLA backbone with a lightweight transposed-convolutional probabilistic action decoder that captures spatial uncertainty through local correlations among neighboring action bins. It further introduces Soft-DTW alignment supervision to accommodate local timing variations between predicted and demonstrated action chunks. Together, these two components reduce sensitivity to demonstrator-specific poses and execution timing, enabling more data-efficient action learning. We evaluate SURE on LIBERO, RoboTwin 2.0, and real-world manipulation tasks. Even with only 10% of the demonstration data or 20% of the training steps, SURE achieves manipulation success rates that match or surpass baseline VLAs.

We consider the Multiscale Single-Index Model (MSIM), first introduced by Oymak and Soltanolkotabi (2021), as a stylized model for hierarchical feature learning with scale separation. Each layer extracts a shared single-index feature at one physical scale and passes it to the next, giving a tractable setting in which to study how deep architectures learn local multiscale representations. Under non-degeneracy and delocalization assumptions on the link function and planted features respectively, for the two hidden-layer case ($K=2$), we analyze joint online spherical SGD under MSE loss, which is a canonical regression objective, in discrete time and show that it achieves weak recovery from random initialization in $O(d\log^2 d)$ samples, followed by strong recovery to accuracy $\varepsilon$ in $O(d\log(1/\varepsilon))$ additional samples. The main technical challenge arises from the non-Gaussian learned features observed by the second layer. We address this challenge through a quantitative Gaussian comparison using delocalization, combined with sharp trajectory-level control of the coupled SGD dynamics. This work is a companion to Bruna (2026, arXiv:2607.03347), which studies recovery under a correlation objective for general $K$.

Particle image velocimetry (PIV) reconstructs a flow field from the motion of tracer particles but the reconstruction quality is bounded by how well the flow structure of interest is resolved in the images. The acquisition position is typically fixed by the operator before the measurement. We cast camera placement as a sequential decision-making problem and train a reinforcement learning agent to control a 6-DoF manipulator equipped with a camera, enabling it to locate, frame and resolve a hydrofoil wake in a water tunnel. The agent commands the camera pose and observes the scene only through the velocity field and uncertainty estimated from time-separated image pairs. It is trained entirely in a photorealistic simulation of our water tunnel, with wakes generated by Computational Fluid Dynamics (CFD) simulations of hydrofoils with varying profiles, positions and angles of attack. After zero-shot transfer to the real water tunnel, the policy reliably locates and frames the wakes of randomly placed hydrofoils. It resolves the wake at a higher resolution than a fixed camera view and with lower error in the measured flow statistics, in roughly half the time of a full scan and with far fewer image pairs.

Robots increasingly choose actions by searching against a reward model learned from human preferences, in model-predictive control or best-of-$N$ selection; we measure what this does to the true objective. On HalfCheetah and Hopper with crowd-sourced labels, and on Meta-World drawer-open and plate-slide with synthetic labels and blind labels from the two authors, harder MPPI search cuts true return by up to 76%: with human labels the drawer opens in 0 of 20 runs against 10 of 10 for the cloned policy the planner starts from, while held-out label accuracy stays at 0.75-0.97. On plate-slide human labels are exploited less than synthetic ones, and on window-open search helps. True return roughly follows how far the planner's weights move from its proposals ($\sqrt{\mathrm{KL}}$) for both weighting rules we test, and best-of-$N$ with a verifier from synthetic labels falls below a single noisy action. We prove that $χ^2$-regularized MPPI at a temperature set by the scorer's error spread on the proposals is worst-case optimal, that no positive gain can be certified once that spread exceeds the scorer's own, and that an episode's return splits exactly into per-step terms, of which the guarantee covers the selection terms. Calibrated from 100 proposal pairs ranked by true return (an oracle) on start states the evaluation never uses, this rule is the safest fix we test: with each of five independent sets it succeeds in every arm run and keeps on average 86-94% of the cloned policy's return on plate-slide and drawer-open, though it is cautious where search helps. Calibration from a person's comparisons cannot detect a bias the scorer learned from that person; correcting for the pairs people cannot tell apart, or a second person, limits the damage.

Measurement exploration reveals how interacting local controllers behave but can alter the steady-state regime a network operator seeks to optimize. Exploratory rewards may therefore fail to identify a configuration's no-exploration value: under full exploration, identical logged distributions can conceal different target values. Identification requires accounting for both action-law mismatch and shifts in the steady-state context distribution. We introduce CAUSAL-UCB, which selects configurations using optimistic bounds on target causal values. Its interference-aware clipped doubly robust evaluator produces predecision certificates separating statistical variation from calibrated identification, transport, and temporal-dependence errors. For any finite horizon and predetermined schedule of measurement-exploration rates $\{η_t\}$, we prove lower bounds of $Ω((1+\sum_tη_t)^{-1/2})$ on minimax inference error and $Ω(\sum_tη_t)$ on measurement-induced deployment loss in a stationary, potentially slow-mixing family. We bound regret under valid certificates and selected-information growth, and establish an anytime finite-sample certificate with a shared radius for estimation, selection, and reporting in a restricted known-local-interference model. In a high-mismatch benchmark with 24 paired seeds and 200 rounds, CAUSAL-UCB attains target regret of $1.732$ versus $6.326$ for certified IPS-UCB, with deployment gaps of $30.607$ and $23.217$, respectively. Independent-audit coverage tests and trace-driven semi-synthetic studies further assess certificate coverage and deployment trade-offs.

How does using less conversational history change reader-token use and benchmark scores? We examine this question on 1,100 BEAM questions. We held the retrieval ranking and reader constant across five fixed passage budgets, and mean assigned scores rose as the passage budget grew. The score gain per additional token became smaller at the upper tested budgets. Answerable questions generally received higher scores with more context. Pooled abstention scores sometimes rose as the passage budget shrank. We also tested message retention in a controlled task with explicit conversational-state targets and compared retained user messages with token-matched recency. Qwen matched the full state more often for direct requirements when user messages were retained. Its full-state matching declined when answers depended on references to earlier offers. Mistral produced near-floor results across the retention conditions. A scoring audit leaves the correctness of BEAM answers unresolved. These experiments characterize the token and score consequences of the tested memory budgets. Reliable budget selection depends on retaining the evidence each question requires and validating the measure used to assess answer quality.

Graph Retrieval-Augmented Generation (GraphRAG) connects information across documents through entity relationships to support relational reasoning. However, black-box access to GraphRAG systems can enable adversaries to reconstruct knowledge graphs and recover both structural connectivity between entities and the semantics of their relationships. Existing defenses lack formal privacy guarantees and provide limited protection against such reconstruction attacks. We propose RelGuard, the first differentially private framework for GraphRAG, which provides formal privacy guarantees by mitigating connectivity leakage and protecting relations and their semantic meaning through relation removal. We formalize relation differential privacy (relation-DP), which bounds the distinguishability of mechanism outputs according to the semantic distance between alternative relations while accounting for relation removal. Building on this definition, we develop a four-step defense mechanism that sanitizes source documents before GraphRAG indexing by perturbing relation predicates or removing their associated connectivity using the exponential mechanism. Experiments across three GraphRAG systems and two targeted reconstruction attacks demonstrate that RelGuard provides stronger protection against reconstruction attacks than four existing defense mechanisms while better preserving downstream utility.

Parameter-efficient fine-tuning (PEFT) has been increasingly adopted to fine-tune pretrained models with low computational cost and is therefore also preferred in the federated learning paradigm to reduce computational and communication overhead. In this work, we investigate the sparsely explored problem of the risk of private data leakage in LoRA-adapted Vision Transformers (ViTs), where low-rank updates weaken the gradient signal available to a server with access to individual client updates. We show that quality of reconstruction is limited by two coupled ambiguities: reduced information in LoRA gradients and missing spatial cues from frozen positional embeddings. To address this, we propose a spatially anchored generative inversion framework that incorporates server-prescribed trainable positional embeddings into ViT blocks during fine-tuning at clients. The resulting positional gradients act as layer-wise spatial anchors, providing additional positional cues, while the generative prior keeps reconstructions on a realistic image manifold, reducing duplicate-data and patch-permutation artifacts, {which produce similar gradients as those produced when the target client images are used}. Adding our proposed spatial anchoring to GIAS improves PSNR by 10.14 dB, SSIM by 0.45, and LPIPS by 0.098 on CIFAR-10, as well as PSNR gains of 7.83 and 11.74 dB on high-resolution CelebA and CARS datasets, respectively, in the FedSGD setting. Such gains in reconstructed image quality metrics are also maintained at higher batch sizes and in FedAvg settings. These results indicate that LoRA-based PEFT does not, by itself, prevent data leakage in FL, and carefully incorporated positional embeddings can help in exposing private information.

Mutation testing for deep learning (DL) faces a cost-realism trade-off. The pre-training approach produces realistic but computationally expensive mutants, whereas existing post-training operators modify trained models efficiently but produce mutants that less closely resemble real faults. This study asks whether the trade-off is necessary: can the weight-space effect of a fault-informed pre-training mutation operator be predicted from a trained model, without retraining it? We answer this question with TensorMutation, which represents a trained model as a neural graph and uses a mutation-conditioned graph neural network (GNN) to predict parameter changes. The GNN is trained offline and reused across unseen models so that TensorMutation can generate executable mutants without retraining. We construct a corpus of 25,252 original-mutant pairs from 71 multilayer perceptron (MLP) classifiers using 11 fault-informed pre-training operators, including 25 newly curated subjects. TensorMutation partially reproduces the behavioral effects of pre-training mutations on nine held-out subjects, and direct weight-space supervision significantly improves weight-space fidelity. On 12 real-fault subjects, TensorMutation achieves the highest median behavioral similarity among the evaluated approaches. The improvement over the post-training tool DeepMutation++ is statistically significant (Holm-adjusted p=0.018), whereas the differences from the pre-training tool DeepCrime and the norm-matched random control are not significant. These results show that fault-informed pre-training mutations leave learnable structure in weight space that transfers to unseen models.

Antarctic whale vocalization detection requires temporally localizing call boundaries in long underwater recordings and classifying them into predefined call types. Training reliable machine learning models for this task is difficult due to the variability and sparsity of whale calls in passive acoustic monitoring datasets. These difficulties are further amplified by time-varying noise that can mask whale calls and by site-specific ambient sound characteristics to which a trained model may not generalize. In this work, we present NAVE, a Conformer-based detector that improves whale-call detection via three components. First, its input representation combines per-bin complex mean subtraction, which suppresses constant site-specific ambient sound characteristics, with an additional per-channel energy normalization (PCEN) channel that adapts to temporal variations in the noise floor. Second, a frequency-dynamic stem processes these noise-aware representations with convolutions that adapt across the characteristic frequency bands of the target call types. Third, we use a Conformer backbone with a convolution kernel matched to the duration of the target calls that adds an inductive bias and improves temporal localization. We evaluate NAVE on BioDCASE 2026 Task 2 using leave-one-site-out cross-validation across the validation sites, which approximates deployment on unseen recording sites. A single NAVE model with a tuned per-class threshold exceeds the stronger WhaleVAD-BPN baseline, using no boundary-proposal network and no search over post-processing parameters. A controlled ablation study shows that the normalized front-end and a duration-matched convolutional kernel each contribute substantially to recovering the D downsweep, the hardest class for previous detectors, with the wide kernel giving the largest aggregate gain.

Language model deployment often requires different trade-offs between competing objectives as user preferences change, yet standard post-training methods typically produce a single policy for a fixed objective weighting. We introduce LambdaMix, a new parameter efficient method for learning a preference conditioned family of policies with group-relative online reinforcement learning. LambdaMix jointly trains one low-rank adapter per objective over a frozen base model. One preference vector selects the adapter mixture that generates responses and weights the separately normalized rewards used to update the adapters. A single training run supports changing preferences at inference without further optimization. We experiment with four models on helpfulness and harmlessness and on solving math problems while following instructions. We also evaluate the two largest models on rewriting comments to preserve wording while reducing clues about personal attributes. On the first two tasks, LambdaMix achieves higher hypervolume than all evaluated conditioned and policy merging baselines across all four models. LambdaMix uses fewer training completions and stores fewer adapter parameters than a set of GRPO specialist policies. With architecture and initialization fixed, varying adapter mixtures and objective weights together gives wider score ranges and higher coverage than varying either alone.

Model-based reinforcement learning (MBRL) approaches have demonstrated strong performance across diverse domains in recent years, learning accurate world models from agent trajectories. However, despite this progress, current approaches still struggle to surpass human performance on tasks that require long-term memory and planning. In this work, we address the problem of long-term memory in MBRL by moving beyond the commonly studied paradigm of architecture design. We introduce Test-Time Model-Based Reinforcement Learning (TT-MBRL), using gradient descent at test time to encode the memory of episode observations into world model parameters. This enables the agent to learn a distinct world model for each episode, improving memory capabilities compared to traditional approaches that solely rely on contextual memory. We first establish a proof of concept on a simplified two-dimensional variant of the Memory Maze environment, which we refer to as Memory Maze 2D, showing consistent improvements across different world model architectures. We then demonstrate the effectiveness of TT-MBRL in the 3D domain, achieving new state-of-the-art results on the Memory Maze benchmark. Finally, we show that TT-MBRL maintains strong performance on non-memory benchmarks such as DeepMind Control and Atari 100k.

Trial-and-error FE simulation of sheet forming could be computationally expensive, neural-network-based learned simulators are used to accelerate design. These simulators should accurately capture localised responses, such as thinning around sharp radii or near ribs, which determine formability. Transformer-based neural operators propagate information efficiently across the mesh but tend to smooth such localised responses. This study develops BiLoGo, a Bipartite Local-Global learned simulator with an FE-guided design that connects local node-element message passing with a Transformer-based neural operator for long-range information exchange. BiLoGo has three modules: a bidirectional node-element GraphSAGE model that learns nodal-elemental coupling on a bipartite graph and predicts each quantity on its native FE entity without interpolation; a dual-stream Transformer-based neural operator that processes blank deformation and tool contact with separate token sets, preventing localised contact information from being averaged with the whole-blank state; and a fusion and element update module that combines global and local features via gated residual fusion and updates element features conditioned on element-node kinematics. Evaluated on a dome-shaped cold forming case and two corner-shaped hot stamping cases (one with ribs), BiLoGo achieves the lowest positional and threshold-based thinning errors on all three cases compared with single-level and hierarchical graph-based models and Transformer-based neural operators. Near the ribs, the smoothing index of predicted thinning increases from 0.41-0.56 for Transformer-based neural operators to 0.94 for BiLoGo-U. Compared with the hierarchical graph network CAtt-BiUGNN, BiLoGo-T reduces mean positional error by 15.5% on the Corner case and BiLoGo-U by 8.3% on the Corner+Rib case, with 66-69% less training memory.

We present a unified potential method for contextual search and bilateral trade with one-bit feedback and obtain order-tight algorithms for both problems. Our framework starts from a Bayesian perspective in which the unknown target (for contextual search) and the unknown pair of valuations (for bilateral trade) are sampled from a known prior distribution. In that model, the decision maker bases its prediction on the posterior distribution, obtained by conditioning on feedback received in previous rounds. For Bayesian contextual search, we present an algorithm with a total loss bounded by $O(\mathrm{Tr} \sqrt Σ)$, where $Σ$ is the covariance matrix of the target's prior; and for contextual bilateral trade, our algorithm's regret is bounded by $O(\mathrm{Tr} \sqrt{Σ_S}+\mathrm{Tr} \sqrt{Σ_B})$, where $Σ_S$ and $Σ_B$ are the covariance matrices of the seller's and buyer's valuations. We then use ideas from this Bayesian setting and adapt a ``dual'' of this trace-square-root potential to the classical worst-case setting, where the target and valuations are chosen adversarially from the unit Euclidean ball. For both contextual search and bilateral trade, we obtain tight $O(\sqrt d)$-regret algorithms. These results match the existing lower bounds up to a constant factor and improve on the previous guarantees of $O(d\log d)$ for contextual search and of $O(d6^d)$ for bilateral trade.

Large Language Model training is increasingly constrained by memory capacity. Activation Checkpointing reduces activation memory by avoiding storing selected activations during the forward pass and rematerializing them during backpropagation, thereby trading memory savings for additional computation. Although this technique is essential for scaling model training, rematerialization can be prohibitively expensive, particularly for Transformer linear layers. Approximate computation could reduce this overhead, but conventional approaches either introduce errors that accumulate over the backward pass or fail to exploit information available from the initial forward pass. We introduce a novel approximate rematerialization method that can yield unbiased gradient estimates: using information collected during the initial forward pass, the method selectively recomputes only the most important layer components, either through greedy selection or randomized sketching. Ablation studies show that the resulting gradients closely match the exact gradients. When applied to feed-forward layers, attention layers, or both, our method can recompute only two-thirds of the layer components in a Llama-488M model, while increasing perplexity by just 1.7%, 1.2%, and 2.5%, respectively.

We study kernel-based learning of analytic operators from finite samples, with a focus on how the regularity of the target should determine the choice of the kernel. We establish embeddings of analytic operators into vector-valued reproducing kernel Hilbert spaces and derive finite-sample, high-probability guarantees that quantify the uncertainty due to random sampling: in the noiseless setting, suitable analytic kernels yield rates exponential in a fractional power of the sample size, while Sobolev-type kernels yield polynomial rates. We apply the theory to the solution operator of a parametric elliptic PDE, where the training data are themselves finite-element approximations, and obtain bounds that separate the sampling and discretization contributions to the error and prescribe how the mesh size should be coupled to the sample size. The results are illustrated numerically.

A common test of personalization replaces a user's history with someone else's and measures the drop in accuracy. But this swap also changes the topics in the history, so the drop may reflect topic mismatch. We examine this problem with a frozen language model on LaMP-2 news categorization, using histories from similar users and synthetic histories assembled from other users' articles. The user's own history improves accuracy by 2.3 percentage points over a model trained without history. Removing history only at evaluation makes this gain appear larger, at 4.8 points. Category labels alone reproduce 86% of the accuracy gap between the user's own history and a random user's history. With synthetic histories matched on labels and text similarity, the gap falls to 0.9 points. Across two user splits and three training seeds, the advantage over a similar user's history is 3.2 points for users seen during training and 0.65 points for unseen users, where the evidence for an advantage remains inconclusive. A small movie-tagging experiment with one training seed also finds no advantage over other users' histories. These findings show why personalization should be evaluated on unseen users, with histories of similar content and a model trained without history as comparisons. We also find that adding gated cross-attention improves on a prefix-only model under our training setup; whether this holds more broadly remains open.

Auditing outsourced LLM fine-tuning requires checking an untrusted provider's training while controlling verification cost. Uniform sampling ignores training-time signals, while checking a suspicious record can require replaying several preceding steps. We present ReplayAudit, a telemetry-guided auditing system that seals training records before challenges and independently replays sampled intervals from sealed checkpoints. Loss changes and parameter-update norms guide stratified random sampling; replay determines whether the selected computation agrees with its records. We relate sampling probabilities to replay work and evaluate the system on real LoRA training. A 7B BF16 prototype rejects a false update-norm report during a random audit; targeted checks also reject false loss reports and skipped updates. We calibrate sampling budgets offline to at least a 90% probability of directly selecting an injected step. For two learning-rate deviation settings, executing these budgets on the same honest 1.5B trajectory reduces median complete-audit wall time by 49% and 67% relative to uniform sampling. When skipped updates evade the high-anomaly stratum, the corresponding audit instead costs 38% more. These results show how deviation capture and actual replay work jointly determine the cost of telemetry-guided auditing.

We study coresets for clustering problems, focusing on the classical $k$-median and $k$-means objectives. A coreset is a small weighted summary of the input data that approximately preserves the clustering cost for every choice of centers. Such summaries are a central tool for scaling clustering algorithms to large datasets. In this work, we consider the capacitated setting, where the clustering is subject not only to choosing centers but also to respecting prescribed cluster-size constraints. This model is a natural generalization of standard clustering and has been widely studied, but its coreset complexity is less well understood. For finite metric spaces, we obtain optimal-size coresets; surprisingly, these match the best-known bounds for unconstrained clustering. We also give improved coreset bounds in Euclidean space, notably the first coresets with near-linear dependence in $k$ while being independent of the dimension and input size. The main new technical ingredients are concentration bounds for the dual solutions of min-cost flow problems, which allow for both significantly simpler and tighter analyses of the underlying sampling distribution used in the coreset construction.

Private, personalised language models must run on local consumer devices, where large, frontier models cannot fit, and retaining long contexts is non-trivial. Sequential memory---the key-value (KV) cache---grows linearly with context length, quickly dwarfing the model itself: for Qwen3-1.7B-Base, 200,000 retained entries require 21 GiB, 6.6 times more than the model weights. Yet the KV-cache compression methods we measured fail beyond moderate compression ratios. We introduce Transformer-FSQ, addressing extreme KV-cache compression by replacing standard key and value projections with a finite scalar quantisation bottleneck. Our cache stores a packed vector of discrete integer codes per token, with no per-token quantisation metadata. Converting a pretrained transformer trains only these projections on 0.25 tokens per original parameter. We evaluate Transformer-FSQ against five baseline families on models from 0.6B to 8B parameters. On long-context retrieval (RULER), from 12$\times$ to 97.5$\times$ Transformer-FSQ is the most accurate method we evaluate, across scales. A fused Triton kernel decodes directly from the integer codes; with baselines given their authors' kernels where available, it decodes at $1.01\times$ (95% interval $0.98$--$1.05$) the throughput of native SDPA attention on the uncompressed Qwen3-8B-Base at $32\times$ and $1.54\times$ at $97.5\times$. Extreme KV-cache compression is therefore a practical way to extend how much context a fixed device can hold.

Rubric-based reinforcement learning (RL) extends RL with verifiable rewards to open-ended tasks by scoring each response against query-specific rubrics. Under group-relative optimization, however, a rubric that every sampled response fails (unreachable) or passes (saturated) gives all responses the same outcome and no learning signal; in Qwen3-4B medical training, 60.88% of query-rubric observations are inactive. We find that the two states lack different supervision: stating an unreachable rubric in the prompt raises its satisfaction from 1.13% to 91.25%, whereas responses that all pass a saturated rubric still differ in quality. We therefore propose RuACT (Rubrics as Adaptive Critics and Teachers), which assigns each inactive rubric the role that supplies what it lacks. Rubric as Teacher uses unreachable rubrics as privileged information for reward-gated on-policy self-distillation, and Rubric as Critic decomposes saturated rubrics into child rubrics that separate current responses while sharing the parent's weight. RuACT outperforms GRPO, rubric reweighting, and rubric-guided exploration on all four medical and scientific benchmarks with Qwen3-4B and achieves the best average with Qwen3-1.7B. On initially unreachable rubrics, unguided recovery rises from 35.02% to 42.86%, while child refinement improves reward-quality alignment on independent rollouts from 57.74% to 60.77%. Code is available at https://github.com/ZJU-REAL/RuACT.

Adam estimates second moments coordinatewise, but typically adds the same scalar $\varepsilon$ to every denominator. We study how this shared $\varepsilon$ affects convergence when noise and curvature vary across coordinates. Under coordinatewise finite-$p$ noise moments, $p\in(1,2]$, choosing $\varepsilon_i$ from the noise and curvature bounds for each coordinate attains the heavy-tailed $\ell_1$-stationarity benchmark $O(T^{-\frac{p-1}{3p-2}})$, matching the known optimal time exponent for smooth nonconvex stochastic optimization. For $1<p<2$, Adam with a shared $\varepsilon_t$ admits a lower bound of order $Ω(T^{-\frac{p-1}{2p}})$. This lower bound holds for every predictable scalar choice of $η_t$, $β_{1,t}$, $β_{2,t}$, and $\varepsilon_t$, precluding a dimension-uniform benchmark rate for this class.

AI weather prediction models are released with 32-bit (fp32) weights and run in fp32 by default, although current GPUs compute much faster in 16-bit formats. We run five publicly available models -- Aurora, Pangu-Weather, FourCastNet, SFNO and FastNet -- on a single 16 GB GPU from 52 initial dates in 2024, in fp32 and with their networks in half precision (fp16), and verify the forecasts against ERA5. The physical state must stay in fp32: converting the whole model rounds it in physical units, which in fp16 overflows and in bf16 discards meaningful information. With the state, its normalisation and denormalisation in fp32, running the network in fp16 changes the global RMSE by at most 1.4%, and by about 0.5% or less for four of the five models, equivalent to less than one hour of lead time up to day 8 and at most two hours at day 10. The differences show no systematic pattern by region, scale or for tropical and extratropical storms, apart from a small tropical penalty in FourCastNet, and fp16 leaves the CRPS, spread and calibration of Aurora's ensemble almost unchanged. No value overflowed fp16's numeric range (+/-65,504); the largest stay 38 to 544 times below it. A 10-day forecast in fp16 is 1.3-3.5 times faster than in fp32 and uses 1.4-3.7 times less energy. Sensitivity experiments show that rules of thumb from other applications of reduced precision carry over only partly: the first and last layers dominate for some models but not for others, and no type of layer is sensitive on its own. For running AI weather models on a single GPU, fp16 with the state in fp32 is practically lossless at a fraction of the cost. bf16, often preferred for its wider range, is as fast but less accurate than fp16 for every model, and simple 8-bit quantisation costs far more skill.

Post-training language models for reasoning typically rely on gold answers, task-specific verifiers, or learned reward models, limiting scalability when such signals are scarce. Multi-agent debate offers a self-supervised alternative, but converting debate into reliable training supervision remains challenging: majority voting weights agreeing responses equally, lacks a principled tie-breaking rule, and is ill-defined for open-ended tasks without a canonical answer; likelihood-based scoring can conflate fluency with correctness; and homogeneous committees may reinforce correlated errors. We introduce Alignment by Confident Debate (AbCD), which converts debate traces into preference data without gold answer labels. A sampling-diverse committee debates each problem, after which final answers are clustered by task-level semantic equivalence. We use verifiable tasks (mathematical reasoning, multiple-choice quantitative QA, and executable code-generation) as a controlled test bed: their answer keys or executable checkers let us evaluate the quality of debate-derived supervision against ground truth, even though the oracle itself never uses gold labels. Post-training on the resulting preferences outperforms inference-only ensembling and debate, as well as SFT on the same traces, across five model families spanning 1B to 8B parameters and four datasets, improving consensus accuracy, with a maximum gain of $21.5_{pp}$. Ablations show that the approach transfers to a larger model and to code generation, while also revealing limitations of heterogeneous committees and agreement-based weighting. Overall, the results suggest that debate traces can yield useful preference supervision when answer equivalence is combined with confidence-weighted disagreement.

Machine-learning surrogates are increasingly used to accelerate computational fluid dynamics, yet progress is limited by the lack of benchmarks capturing realistic, time-dependent turbulent flows. We introduce TurtLES, a 14 TB dataset of 3D turbulent cylinder wakes in the shear-layer transition regime, computed via high-fidelity implicit large-eddy simulations. The dataset contains 400 trajectories of 400 time steps each, with 3-11 million points per frame, spanning variations in geometry, Reynolds number, and angle of attack. Unlike existing datasets, TurtLES combines (i) physically realistic turbulence with an active energy cascade driven by vortex stretching, (ii) dense irregular meshes, and (iii) long-horizon temporal dynamics. We benchmark state-of-the-art neural operators on autoregressive rollouts and identify key failure modes, including temporal instability, loss of small-scale energy, and challenges in learning on large irregular domains. TurtLES provides a new benchmark for evaluating surrogate models on 3D turbulent wakes at industrially relevant mesh resolution.

Decision-focused learning trains prediction models to improve downstream decisions, but many methods require costly solver calls or optimal-solution labels. We propose a solver-free decision-focused learning framework for NP-hard edge-selection problems with uncertain parameters in the objective, the constraints, or both. A neural optimization layer is first trained without solution labels, using problem objectives and constraint penalties. Graph-informed edge scoring and iterative refinement encourage concentrated and structurally consistent continuous decisions. The layer is then frozen, and a prediction model is trained through its continuous output using a problem-specific loss evaluated under observed parameters. At deployment, the predictor, optimization layer, and a lightweight decoder produce decisions without solving the original NP-hard problem. Experiments on the asymmetric traveling salesperson and endpoint-free orienteering problems show consistently better decision quality than prediction-focused learning and improvements over most evaluated decision-focused baselines. Our framework requires less training time than the evaluated decision-focused baselines, even when optimization-layer pretraining is included. Neural deployment also produces decisions faster than solver-based deployment, offering a better quality-efficiency trade-off. The layer also reduces training cost in existing decision-focused methods when used as an approximate optimization oracle.

Learning to profit from trade requires learning both sides of a market: what a seller is willing to accept and what a buyer is willing to pay. We study repeated bilateral trade, where an intermediary posts a purchase price to a seller and a sale price to a buyer whose valuations are drawn from an unknown joint distribution. Trade occurs only if both parties accept their respective prices, in which case the intermediary earns the difference between the two prices. The intermediary observes only whether trade occurs, without observing either valuation. We show that learning both sides of the market is no harder, in terms of regret, than learning just one. Specifically, we design an algorithm achieving regret $\widetilde{\mathcal{O}}(T^{2/3})$ against the best fixed pair of prices, matching, up to logarithmic factors, the classical lower bound for one-dimensional dynamic pricing. Our result shows that the classical one-dimensional regret rate survives the transition to two dimensions, raising the question of whether this absence of a curse of dimensionality extends to arbitrarily high dimensions.

Long-context language models often neglect information in the middle of the input, a phenomenon known as "lost-in-the-middle". Part of this gap may stem from the long-range decay of rotary position embedding (RoPE), which weakens the positional signal between distant tokens. Existing input compression deletes low-importance tokens, and attention reweighting acts only after RoPE has weakened this signal. RoPE-level methods modify positions directly but apply the same remapping to every query. To address these limitations, we propose QGPM (Query-Guided Positional Modulation), a training-free method that reshapes RoPE positions using query relevance from the model's own attention heads. QGPM builds a broad relevance map for lower layers and a sharp one for upper layers. Low-relevance tokens advance by less than one position on fast-rotating planes, pulling relevant evidence toward the query. Slow-rotating planes keep the original layout, with every token in place. Experiments on NoLiMa and RULER with multiple context lengths show that QGPM improves over base models and strong baselines, with the largest gains on retrieval tasks, and generalizes across multiple backbones. Our code is available at https://github.com/iZHpy/QGPM.

Agentic AI increasingly enables software development activities to proceed without continuous human involvement. While this autonomy may accelerate development, it also raises questions about when developers can meaningfully oversee software development done by agents. We propose deliberate pacing: intentionally adjusting when and how autonomous development proceeds to preserve opportunities for meaningful oversight. Drawing on an exploratory multivocal literature review of 20 sources, we identify three factors motivating this vision: changing relative costs of human involvement, the timing of oversight opportunities, and interdependencies between pacing decisions across the software development life cycle. We argue that neither uninterrupted autonomy nor frequent human intervention is universal. Instead, pacing should balance opportunities to understand and influence consequential decisions against costs in human attention, development time, and rework. We outline a research agenda to investigate when deliberate pacing is beneficial, how intervention timing affects oversight, and how pacing decisions interact across development activities.

Protein conformational ensemble generation aims to efficiently characterize the structural variations accessible to a protein beyond a single static structure. Recent flow-based methods efficiently generate conformational ensembles, but they model coordinated structural variation implicitly, which can limit ensemble fidelity and hinder direct control. To address this limitation, we propose CoMoFlow, a controllable framework for generating protein backbone conformational ensembles by explicitly modeling collective variation. CoMoFlow constructs protein-specific collective modes from the equilibrium structure, learns the distribution of their coefficients across the molecular dynamics (MD) ensemble, and uses sampled collective variations to guide SE(3) flow matching through feature conditioning and prior recentering. Experiments on ATLAS and de novo proteins show that CoMoFlow improves reconstruction of MD conformational distributions and coordinated structural variation relative to strong flow-based baselines, while enabling direct control over targeted conformational changes.

We study the problem of multiclass classification within the framework of marginal-nonuniform learning, where the constants in the learning rate may depend on the unknown marginal distribution generating the unlabeled data, but not on the target concept. This notion naturally lies between classical PAC learning, where the constants are independent of both the marginal distribution and the target concept, and more recent universal learning, where the constants may depend on both. We give a complete characterization of the optimal marginal-nonuniform learning rates, establishing a fundamental trichotomy: exponential, linear, and arbitrarily slow rates. This characterization is witnessed by a new combinatorial dimension, termed the uncentered DS-eluder dimension. A key ingredient in the upper-bound proofs is a set of new distributional exhaustion lemmas that exploit the sequential nature of this dimension and may be of independent interest.

Traditional information theory provides powerful tools for analyzing statistical properties of information but offers limited insight into meaning and semantic transformation. It has never been possible to quantify the meaning of an arbitrary piece of information. Large language models represent information as vectors embedded within high-dimensional semantic spaces, enabling meaning to be analyzed geometrically. This paper defines semantic state as the vector representation of information within a model's semantic space and defines cognitive work as the transformation between semantic states, movement through semantic space. To quantify this transformation, we introduce the Contextual Information Differential (CID), a covariance-aware metric based on Mahalanobis distance measuring semantic displacement relative to the structure of the embedding manifold. By treating meaning as a geometric property of semantic space, CID provides a practical method for measuring semantic change, cognitive transformation, and information processing. This framework establishes a foundation for analyzing cognitive processing, learning, reasoning, knowledge generation, synthetic expertise, and human/AI collaboration through implementation-independent measurement of semantic state change in both human and artificial systems.

The task of weakly supervised video anomaly detection is challenging. Existing approaches mainly model video frame features and determine whether anomalies are present based on their magnitudes. In this work, we propose a novel method that incorporates external knowledge into training to improve model performance, robustness, and generalization. We leverage textual prompts and a large language model (LLM) to establish prior information for the specific task, and employ dictionary learning to normalize and compress this information. Moreover, during experiments, we observed inconsistencies in the textual descriptions extracted from similar video frames by image caption models. To address this, we devise a method that leverages an LLM and auxiliary information to ensure consistency in the textual descriptions of similar frames within a video. Finally, we introduce a feature processing approach that incorporates prior knowledge and cross-modal attention mechanisms. This approach learns the relationships and characteristics among different modalities to construct more accurate and reliable feature representations, thereby improving the accuracy and stability of downstream tasks. Experimental results demonstrate that our method surpasses other state-of-the-art approaches.

The mechanical response of a crystal is governed by the evolving network of dislocations it contains, yet quantitatively linking the `state of that network' to its yield stress remains challenging. Here we represent the dislocation networks obtained from discrete dislocation dynamic simulation as graphs, using five distinct graph representations, characterize each network through a compact set of topological properties, and predict the yield stress from these graph properties and the applied strain rate using a standard random forest model evaluated with strain rate stratified cross-validation. The model predicts the yield stress with $R^2\approx0.86-0.90$ across all box sizes and $R^2\approx0.91-0.94$ for larger box sizes, and this accuracy is consistent across all five representations, indicating that the topology-stress relationship is a robust feature of the graph properties rather than an artefact of any single graph representation. We propose graph topology as a compact and robust route for connecting the microstructural state of a dislocation network to its mechanical response in addition to established variables.

Bolt preload is a decisive functional quantity in bolted joints because it determines whether external loads are transmitted primarily through frictional clamping or directly through the bolt. Existing preload measurement techniques, including strain-gauge bolts, force washers, ultrasonic methods, and bolt-length measurements, are effective in certain applications but require additional hardware, coupling, calibration, or joint modifications. This limits their scalability for large numbers of standard fasteners and motivates a non-contact optical approach. This paper proposes PreForceVision, a physics-guided optical framework for estimating bolt preload from preload-induced surface deformation of the bolt-head front face. The method treats preload estimation as an inverse deformation-to-force problem, where deformation of the optically accessible bolt-head surface is extracted from reference-loaded image pairs and related to the internal preload state. Preliminary results for an M16 strength class 8.8 bolt show an approximately \SI{12}{\micro\meter} simulated head-face depression at \SI{90}{\kilo\newton}, strain-gauge measurements of comparable magnitude, and an optical 3D measured central depression of approximately \SIrange{10}{20}{\micro\meter}. The contribution is a physics-guided optical methodology combining controlled image-pair acquisition, DIC- or optical-flow-based deformation extraction, multimodal validation, and uncertainty-aware regression for rough preload estimation under controlled conditions.

Quantitative equity selection and portfolio construction in emerging and cross-border markets frequently suffer from computational bottlenecks, remote API fragility, survivorship bias, and rigid execution constraints such as discrete board lots and investor-mandated drawdown limits. In this paper, we present StockASkill, a local-first, full-data-driven AgentSkill engineered as a specialized Quant Co-Processor for autonomous AI agents and quantitative researchers. Tailored for agile portfolios (10^5 ~ 10^7 RMB) where discrete lot execution and capital protection are mandatory, StockASkill embeds a columnar DuckDB OLAP engine over 42.44 million historical price records and 186,432 financial disclosures spanning 26 years across China A-shares, Hong Kong, US equities, and ETFs. By eliminating ad-hoc network dependencies during analysis, StockASkill reduces cross-sectional screening latency across 5,400+ equities from over 25 seconds to sub-0.5 seconds (a 95%+ reduction). We formalize an end-to-end mathematical pipeline integrating a 7+1 factor alpha model, an SLSQP convex optimizer under smooth CVaR drawdown surrogates with exact path projection, dynamic 100-share board lot integer quantization, and an autonomous thesis lifecycle finite-state machine. Furthermore, we provide a zero-disk, strongly-typed JSON streaming interface designed for high-throughput LLM tool-calling. Open-sourced at https://github.com/axjing/stockaskill, comprehensive empirical benchmarks from 2018 to 2024 demonstrate that the architecture achieves annualized Sharpe ratios exceeding 1.48 and Calmar ratios up to 2.17, while reducing max drawdown by up to 85% relative to benchmark indices.

Transformer inference repeatedly invokes dense matrix multiplication and attention, including its data-dependent softmax normalization. We present Atik, an open-source RoCC accelerator for RocketChip that executes matrix multiplication and dense attention through one shared fixed-point MAC mesh. Atik accepts and returns BF16 tensors, uses a streaming online-softmax schedule that never materializes the full score matrix, and is invoked through a small descriptor-based ABI. We evaluate three meshes cycle-by-cycle on FireSim and report standalone Sky130 physical-design area. Across 137 logged FireSim test records, all tests pass their benchmark tolerance with zero reported output mismatches. Relative to the same scalar Rocket BF16 reference, the 8x8 configuration reaches 31.5x peak matmul speedup and 225.8x peak attention speedup. Hybrid projections for the single-layer workload-derived C models, using directly measured accelerator-kernel cycles, give aggregate speedups of 44.2x, 77.1x, and 32.3x for ViT, TinyBERT, and GPT-2 prefill, respectively; these are not full-model hardware replays.

In many domains from species identification to cancer diagnosis, labelling data to train predictive models remains costly. In these settings, it is advantageous to intelligently select data points to label to achieve the maximum information gain, this is the motivation behind the study of active learning. In addition to the high costs of acquiring data in such domains, there is the added complexity of data distribution shifts that can occur between training and deployment. For example, the features of cancerous skin cells taken in one country may be different in another region of the world. These two challenges - efficient labelling of data, and learning models robust to distribution shifts have been studied separately but not together, even though these challenges often co-occur in real-life settings. We identify failure modes of standard active learning approaches when the distribution of data at deployment is different from the distribution in which the data is sampled for labelling. We propose a new strategy for efficiently learning classifiers that generalise under specific forms of distribution shift, namely subpopulation shift. Drawing on work in learning from heterogeneous data, we aim to train models that perform well across subpopulations of the data, through active learning. To handle this challenging setting, we exploit the fact that many datasets include metadata about data collection that can serve as a noisy proxy for each sample's underlying subpopulation. We demonstrate how to leverage these noisy labels during active learning selection, and demonstrate empirically that our approach improves classifier accuracy across subpopulations.

Real-time Voice AI agents must separate the active user from intelligible background speech that survives conventional noise suppression. Background sources include television audio, family members, and colleagues. Production denoisers remove non-stationary environmental noise effectively. Background speech shares the spectral characteristics of the target and passes through, causing false acceptances in downstream speaker verification (SV) and automatic speech recognition (ASR). We benchmark open-source speaker embedders TitaNet-Large, TitaNet-Small, ECAPA-TDNN, CAM++, and SpeakerNet under a strict sub-second latency budget. Evaluation combines a controlled benchmark with a weighted human-annotated production set. Our controlled benchmark is built from real calls in which a second real speaker is injected into silence intervals at modulated amplitude. We report three findings. First, denoising before embedding is generally harmful and raises equal error rate (EER) across architectures. Second, truncating audio below the full voice-activity segment degrades every model. Third, roughly half of the TitaNet embedding is redundant for this task. Trailing dimensions encode population attributes such as language and gender. Such attributes carry little discriminative value in a demographically homogeneous user base. We distill these results into deployment guidance. Verification should run on raw audio with full voice-activity chunks. Thresholds should favor precision over F1. Truncated embeddings reduce the cost of similarity search.

Returned products in circular engineering often have heterogeneous degradation histories, making remaining function difficult to assess. For bolted joints, corrosion alters surface morphology, contact, friction, preload retention, and mechanical response. This paper presents a physics-informed multimodal framework for linking microscopic and camera-based degradation observations to functional indicators of corroded bolted joints. The workflow defines the corrosion failure mode, generates synchronized inspection and functional-response records, preprocesses image, roughness, exposure, and joint descriptors, aligns microscopy and camera observations through similarity learning, and constructs a calibrated corrosion-severity field. Partial differential equation (PDE) discovery is applied to this field to identify candidate effective mesoscale laws rather than complete electrochemical mechanisms. The corrosion-evolution model is coupled to a separate functional model, with friction coefficient as the primary target and preload loss included when clamp-force data are available. The contribution is a reproducible pathway for physically constrained assessment of degraded bolted-joint behavior.

While MLLMs can often identify what happens in a video, they still struggle with fine-grained spatio-temporal reasoning, which requires precise grounding when and where the supporting evidence appears and composing such evidence into a coherent reasoning chain. This challenge is exacerbated by the lack of scalable data construction and evaluation pipelines that provide granular, verifiable supervision linking reasoning steps to explicit spatio-temporal evidence.To address this gap, we present V-STORM, a structure-aware framework that transforms videos into Spatio-Temporal Scene Graphs (STSGs) and derives Grounded Chain of Thought (GCoT) from validated evidence paths, making each reasoning step traceable to structured and verifiable spatio-temporal evidence.Under this framework, we construct a 260K-sample training set and a 3K-sample human-verified benchmark, both featuring GCoT annotations that link reasoning to explicit spatio-temporal evidence.Experiments reveal a clear separation between evidence localization and evidence composition: current MLLMs can often identify relevant spatio-temporal cues, yet still fail to integrate them into correct multi-step reasoning, exposing a bottleneck that answer-only evaluation largely obscures. Moreover, fine-tuning with V-STORM substantially improves both answer accuracy and grounding faithfulness, validating its effectiveness as supervision for faithful and evidence-grounded video reasoning. Our dataset, benchmark, and source code will be publicly released.

We study the effectiveness and robustness of open-weight large language models (LLMs) for Danish offensive language detection, an underexplored low-resource task. We evaluate four model families (Llama, Mistral, Gemma, and Qwen) across eleven prompting strategies spanning five design categories, cross-prompt generalization, Danish orthographic variation, and robustness under balanced and naturally imbalanced class distributions. We compare against three BERT-family baselines with distinct pre-training strategies, evaluated via both masked language model prompting and supervised fine-tuning. Even without task-specific fine-tuning, LLMs can surpass all three fine-tuned BERT baselines, and lightweight finetuning yields substantial further gains. No single model or prompt category dominates, underscoring that prompt design interacts with model pre-training in ways that cannot be predicted a priori. Semantically grounded prompts transfer more robustly, while format-constrained patterns consistently fail to generalize. ASCII transliteration of Danish-specific characters (æ, ø, å) generally degrades performance, indicating these characters encode meaningful signal. At the same time, the gap between a balanced test set (best F1 of 0.84) and the corpus's natural class skew (0.62 trained on skewed data, 0.70 trained on balanced data) -- relative drops of 26.2\% and 16.7\% -- shows that class imbalance remains a critical open challenge. Our code and configurations are available at https://github.com/PASHAADC/Danish-LLM-HateSpeech.

Reliable uncertainty estimation is important for deploying evaluation metrics in real-world settings, where both input ambiguity and domain shifts are common. In this work, we propose an uncertainty-aware extension of the BEiTScore image captioning metric, producing conformally calibrated confidence intervals alongside alignment scores. Our approach models predictive distributions over human judgments, enabling both accurate scoring and principled uncertainty quantification in a single forward pass. We evaluate our BEiTScore++ metric across multiple benchmarks with human-annotated ratings, showing that it preserves the ranking correlation of BEiTScore, while producing intervals whose width tracks the prediction error more closely than Monte Carlo dropout on most benchmarks. The estimated uncertainty also grows under image degradation and on out-of-domain artwork images, making BEiTScore++ a practical tool for uncertainty-aware evaluation of image-caption alignment.

Existing time series language models predominantly follow an offline, one-shot paradigm: they process fixed-length inputs independently and do not retain past states. This static approach fundamentally misaligns with the continuous nature of real-world streaming applications and fails to maintain the historical context necessary for cross-window reasoning. To bridge this gap, we introduce Continuous & Live Time Series Understanding (CL-TSU), a novel paradigm that shifts the focus toward maintaining a persistent, stateful understanding over open-ended data streams. To systematically evaluate this paradigm, we present LiveTS-Suite, the first comprehensive data suite tailored for CL-TSU, supporting data generation, model training, and systematic evaluation across 18 tasks spanning six core capabilities. It comprises over 5K streams totaling 2.2B timesteps, annotated with 15K specific response events. We further design an evaluation protocol that jointly assesses response timeliness, content accuracy, and computational efficiency. Building on this foundation, we propose TimeOmni-Live, the first live time series understanding model. It integrates a hybrid time series encoding for natively temporal input, a stream-persistent hybrid memory mechanism to effectively preserve exact historical evidence and long-term task states, and an elastic stream-inference synchronization strategy to reduce processing lag. Evaluations show that TimeOmni-Live establishes a strong baseline for CL-TSU, outperforming frontier LLMs (90.0 vs. 79.3 for Opus 5) with only 4B parameters and achieving 206$\times$ lower p95 latency than its backbone. Deployment on a Jetson edge device further shows its ability to support real-time continuous temporal understanding locally at the edge.

Lloyd's k-means algorithm is a long-running standard for clustering, with applications in data analysis, retrieval, quantization and data curation. In this paper, we introduce an exponential moving average optimization procedure for k-means with mini-batch updates, powered by modern deep learning machinery and new insights. We also establish scaling laws on how decay rate and batch size impact final performance. This gives us a formulaic rule to fix them for multiple problem sizes and embedding datasets, without the need for any hyperparameter sweep. The resulting clustering algorithm, coined e-means, is much faster than Lloyd k-means and converges to lower losses, frequently even beating k-means++. On a wide array of embeddings datasets, it beats the loss of Lloyd k-means at 25 epochs in 1.4 to 2.6 epochs, which translates to 10-45x GPU wall-clock speedups compared to state-of-the-art k-means implementations. As an example, e-means partitions DINO10B, a dataset of 10 billion DINO embeddings, into 16 million clusters in only 2.5 epochs, reaching a loss below Lloyd's at 25 epochs.

Intracranial EEG provides high signal-to-noise ratio and spatial resolution for detecting epileptic activity, but patient-specific electrode placement leads to substantial variation in channel count and montage across subjects. This heterogeneity limits the generalization of existing detectors under subject-independent evaluation. We propose MOSAIC (MOntage-agnostic Self-distilled Adaptive Inter-channel Connectivity learning), a lightweight graph-learning framework for subject-independent iEEG seizure detection. MOSAIC pairs a shared convolutional channel encoder with a patient-agnostic channel-attention module that infers a weighted inter-channel adjacency matrix from the encoded features, requiring no fixed montage graph. An exponential moving average self-distillation objective regularizes each prediction toward a slowly updated per-patient teacher. Across four public iEEG corpora (SWEC-ETHZ long- and short-term, Mayo, FNUSA), MOSAIC outperforms state-of-the-art baselines in F2 and Cohen's $κ$ while learning more stable inter-channel adjacencies, demonstrating that self-distilled connectivity learning transfers across unseen implantations.

Visible-light videos and multivariate diagnostic signals provide complementary observations of tokamak discharges. However, existing fusion pretraining methods, primarily learn representations by reconstructing or predicting observations, leaving the correspondence between these modalities underexplored. Representation learning is further challenged by the fact that plasma-related regions occupy only part of each video frame and diagnostic variables have substantially different sampling rates. We propose FusionCLIP, a multimodal contrastive pretraining framework. Spatiotemporal Stability Selection retains informative visual tokens, while a Temporal Query Encoder integrates asynchronous diagnostic signals using their observation timestamps. Contrastive learning then aligns the two modality representations. We construct EAST Multimodal Diagnostics (EAST-MMD), comprising 1307 EAST discharges, and evaluate the model on three independently fine-tuned tasks: disruption prediction, edge-localized mode (ELM) classification, and $H_{98}$ regression. FusionCLIP achieves an aggregate score of 89.07\%, exceeding the best general-purpose pretraining configuration and an in-domain contrastive pretraining configuration without Spatiotemporal Stability Selection by 6.27 and 2.12 percentage points, respectively. Ablations and attention visualizations further support the contributions of its key components.

Large language models enable fine-tuning-free spoken language understanding (SLU) through in-context learning, but direct query-based retrieval may select semantically similar yet intent irrelevant demonstrations. This paper presents Query-Adaptive Intent Detection via Prototype Extraction (APEX), a two-stage retrieval framework that first selects candidate intents using intent-level prototypes and then retrieves contextual exemplars within the selected intent space. Experiments on the SLURP dataset evaluate joint SLU performance using Exact Match Accuracy (EMA) under full-transcript and perplexity-stratified conditions, with additional retrieval analysis on ASR-generated transcripts. Compared with query-based retrieval, APEX improves EMA by 4.34 and 3.97 percentage points with Qwen3 4B and Gemma-3-4B, respectively, on full transcripts. Under high-perplexity inputs, APEX also yields EMA gains of 7.66 and 2.13 percentage points. Retrieval analysis further indicates that prototype-guided candidate selection achieves higher ground truth intent coverage across multiple input conditions.

Large language models have expanded beyond text to vision, audio, and other modalities, but understanding raw electromagnetic (EM) signals remains challenging. Wideband recordings can contain millions of samples, whereas existing waveform-based EM models are primarily evaluated on short in-phase and quadrature (IQ) segments, leaving long-recording understanding and detailed description insufficiently explored. We introduce LEMMO, a large EM model for instruction-driven recognition and description of long raw IQ recordings. LEMMO combines a linear-attention EM encoder with anti-aliased temporal resampling to process up to eight million complex samples. Its EM-text interaction module conditions learnable queries on instructions before retrieving signal evidence. Compact queries support recognition, while structured queries with auxiliary factual supervision organize scene-level properties, signal attributes, and cross-signal relationships for detailed description. We also construct and release a unified 1.38-TB corpus of real-world and synthetic EM signals, including LEMMO-ShortQA and LEMMO-LongDesc. Compared with MERLIN-LoRA, LEMMO improves mean EM-Bench perception accuracy by 15.77 %, mean strategy-generation ROUGE-L F1 by 4.38 points, and mean LEMMO-ShortQA accuracy by 14.31 %. It also improves categorical and numerical fact F1 on both LEMMO-LongDesc tasks while preserving its language backbone's MMLU performance. Our code and datasets are available at https://github.com/N1rv4n4/LEMMO and https://huggingface.co/datasets/DaftJun/RFLAS, respectively.

Language-model agents use experience through external retrieval and parameter learning. Consolidating stored experience into model weights requires deciding both which experience to train on and when its external records can leave retrieval. Admission based solely on student mastery can exclude useful teacher demonstrations, while removal immediately after training can discard guidance that the model still needs. We introduce ENGRAM, a framework that separates training admission from memory release. A demonstration-aware gate admits skills based on student mastery or sufficient successful teacher demonstrations. The student learns from selected action traces with context from an Engram Trace Graph (ETG). A verification gate then compares the updated model without episodic retrieval against the base model with retrieval on reserved probes. Release requires valid training participation and scores meeting both relative and absolute criteria; records for skills that fail verification remain retrievable. Experiments across interactive and reasoning benchmarks show consistent gains over frozen-weight ETG controls, with ablations supporting demonstration-aware admission and selective adapter activation. A controlled held-out comparison shows that verified release substantially reduces the retrievable footprint while remaining close to retain-all performance and outperforming immediate and budget-matched random release. ENGRAM also improves task performance while reducing input-token usage and model calls. Skill-level verification further reduces retrievable experience through selective archival, retaining external guidance for skills that do not meet the release criteria. Code is available in anonymous.4open.science/r/engram-anonymous.

Smartphones, smart-home hubs, and wearables are increasingly exposed to sophisticated Android malware, while centralized threat-detection pipelines raise privacy and scalability concerns. Federated learning (FL) enables collaborative model training without collecting raw client data, but partial participation and variation in local model quality can destabilize global optimization. In this article, we propose DW-FedADMM, a federated optimization framework that combines primal-dual coordination with performance-dependent client-side weighting. The dynamic coefficient modulates each participating client's local state before the augmented-Lagrangian update, while the dual variables regulate disagreement between local and global model states. We evaluate DW-FedADMM on three Android malware benchmarks, namely Drebin, Malgenome, and TUANDROMD, with 5, 10, and 15 clients holding IID data shards, partial client participation, and communication budgets of 5 and 10 rounds. Each configuration is repeated with ten random seeds. DW-FedADMM obtains the highest seed-averaged accuracy on Drebin (97.64% +/- 0.20%) and the best mean accuracy in 9 of the 18 dataset-client-round configurations, including 7 of the 9 configurations with a 10-round budget. Pooled over all configurations, it reaches 95.42% accuracy, 94.62% F1-score, 96.06% precision, and 99.45% AUC, compared with 95.32%, 94.51%, 95.56%, and 99.41% for FedADMM, while also giving the lowest mean false-positive rate and the smallest client update norm among the six compared methods. Both primal-dual methods are well ahead of the averaging- and proximal-based baselines on Malgenome and TUANDROMD, where the latter often remain at the majority-class solution when clients are fragmented. The gain of dynamic weighting over FedADMM is most pronounced on Drebin and with the longer communication budget, where it also reduces the variation across seeds.

Large language models (LLMs) are increasingly used to annotate social-science texts, but high accuracy does not establish that they apply researcher-supplied category definitions: predictions may instead reflect prior associations with category labels or labeled demonstrations. We introduce a counterfactual audit that reassigns definitions among output labels, distinguishing adherence to supplied definitions from preservation of a dataset's original mapping. Across three political-text datasets, models rarely follow reassigned definitions, and standard fine-tuning strengthens the original mapping without reliably increasing definition adherence. We then study retrieval-augmented annotation. Retrieving the nearest training examples generally outperforms retrieving examples separately within each category despite using fewer demonstrations. On the two tasks with substantial gains, direct \(k\)-nearest-neighbor classification recovers 74--93\% of the improvement, indicating that neighborhood labels provide much of the predictive signal. Relabeling fixed demonstrations redirects predictions, whereas aligning demonstrations with reassigned definitions substantially increases adherence to the new mapping. Thus, accuracy-improving interventions need not increase codebook adherence; definitions, category labels, demonstrations, retrieval, and training data should be aligned and validated jointly.

Agent skills are short documents that teach a language model how to handle a class of tasks. To save context, these documents are increasingly compiled into low-rank adapters (LoRA), which keep the behaviour but lose the text: an adapter is hard to read, edit, audit or move to another model. We ask whether the skill document can be written back from the adapter and its frozen base model, and we score each recovered document by the share of the original's key instructions it states and by the tasks it solves. Reading the weights alone recovered little with the methods we tried; a learned map from weights to text passed on about one bit of the document. Running the adapter worked much better. The Executing Writer, a small writer trained on top of the adapted model, recovers 39-44% of the key instructions of held-out skills, against 1.3-1.5% with another skill's adapter, and on adapters fitted to 41 external skills its documents restore 85-91% of the original's task gain. These strongest results are on adapters whose training stored the document; adapters trained on behaviour alone are much harder for trained writers (1-2% of instructions). For these adapters we test an agent that runs the base and adapted models and writes a skill from where they differ. On ALFWorld its skills recover 58-71% of a LoRA's gain and carry over to models that cannot load the LoRA, lifting Llama-3.1-8B from 5% to 55%. Whether such a skill comes from the adapter depends on what the adapter learned: when its behaviour cannot be guessed, the adapter is the source; when it can, a reward loop without the adapter does as well.

Policy entropy is widely used to diagnose exploration and collapse in reinforcement learning with verifiable rewards (RLVR), yet its magnitude does not consistently track reasoning performance across optimization objectives. We argue that this ambiguity arises because global token entropy conflates uncertainty in successful and failed trajectories, diversity across correct solution strategies, and variation among realizations of the same strategy. We introduce \textbf{Reward-Conditioned Entropy Allocation (RCEA)}, a hierarchical diagnostic framework that separates complete-trajectory entropy from average token entropy, conditions uncertainty on verifier outcome and process validity, measures correct-solution-mode retention, and connects these population-level quantities to local covariance-driven updates. In controlled sweeps over six RLVR objectives at two model scales, entropy trajectories do not induce a stable ordering of reasoning accuracy. Token entropy is higher for verifier-failed than successful trajectories under both broad separately aggregated estimates and a common-question paired analysis, whose confidence intervals exclude equality at 10 of 12 displayed endpoints. A local covariance approximation predicts the direction of adjacent-checkpoint entropy change in 95.5\% of nonfailed transitions; a reward-stratified endpoint proxy further shows that positive chosen-token covariance is more tail-concentrated within failed prefixes. Every reportable trained endpoint exhibits some contraction in independently audited correct-solution-mode breadth, whose positive association with large-\(k\) coverage remains exploratory. These results show that entropy collapse should be diagnosed by which uncertainty is removed, rather than by global entropy alone.

Data-use auditing in Large Language Models (LLMs) aims to determine whether the given data has been used for training, thereby supporting data ownership safeguards. A dominant solution is data watermarking, which embeds traceable signals into training data and detects them in model outputs. Recent studies have proposed many watermarking-based auditing methods in LLMs. However, systematic investigation from an adversarial perspective remains limited: non-compliant LLM providers would use unauthorized copyrighted datasets for training their own models to enhance model capabilities, while attempting to evade detection by auditors, thereby avoiding legal liability. To fill this research gap, we propose a novel evasion framework, termed Logit Laundering, where the non-compliant LLM provider manipulates model outputs at inference time to undermine detectable watermark signals. Specifically, we design a logit reshaping module to suppress watermark-induced statistical signals by integrating a statistical bias estimator as a plug-and-play component. This design keeps the model parameters unchanged and performs manipulation only in the logit space, thereby enabling auditing evasion while minimizing the impact on model's learned knowledge. To further balance evasion effectiveness and model utility, we introduce a confidence-aware scaling module to adaptively calibrate the suppression strength based on next-token prediction confidence. Extensive evaluations show that Logit Laundering can successfully evade existing various state-of-the-art auditing methods, significantly attenuating detectable signals, while preserving model utility on downstream tasks. Our work highlights the need for exploring more robust auditing methods to guard against evasion behaviors.

Ensuring high returns while providing strong safety guarantees in offline imitation learning (IL) is a fundamental challenge: existing approaches either rely on explicitly specified cost functions, which are difficult to design for complex real-world safety constraints, or require expert demonstrations that are both safe and high-performing, which are often too costly to collect. We introduce dual-expert imitation learning, which leverages two complementary data sources: (i) safety-compliant demonstrations that faithfully follow safety guidelines but may be suboptimal, and (ii) performance-oriented demonstrations that achieve high returns but may violate safety guidelines. To address this problem, we propose DexDICE (Dual-expert offline imitation learning via stationary DIstribution Correction Estimation). Instead of conventional distribution constraints, DexDICE combines DICE-style offline learning with safe support constraints, enabling agents to exploit high-return behaviors while remaining within safety-compliant regions. We prove that the ideal solution satisfying this constraint exactly has zero unsafe-region visitation with high probability. We evaluate DexDICE on a re-curated DSRL benchmark and validate it on a real-world mobile-robot navigation task trained directly from human teleoperation demonstrations, where it substantially outperforms baselines and remains robust to data scarcity and quality degradation.

We present MINE (Memory-efficient INference Engine), a lossless entropy coding approach that accelerates single-stream inference of Large Language Models (LLMs), where existing lossless compression methods yield speedups only for specific data types such as \texttt{BF16} and under additional assumptions. Single-stream decoding is strongly memory-bound, since every weight must be streamed from memory once per generated token and requests cannot be batched in local or low-latency deployments. At the core of MINE is dtANS-LLM, a GPU entropy coding scheme adapted from the prior dtANS to the needs of LLM inference. It exploits the non-uniform distribution of weights, which persists even after quantization, and exposes hyper-parameters that trade compression ratio against decoding cost. We auto-tune these parameters for each tensor and GPU to obtain the Pareto front between model size and decoding latency. MINE integrates into vLLM as a plugin and selects the best kernel configuration for a given memory budget. On Llama~3.1~8B~Instruct with a 2.1-bit EntQuant checkpoint, MINE achieves speedups of 2.52$\times$ and 1.48$\times$ over the fastest baseline on an RTX~4060~Ti and an RTX~5090, at effective bitrates of 2.24 and 3.43~bits per weight. MINE is available as open source.

In the healthcare sector, claim denials impose significant financial and administrative burdens, emphasizing the need for automated and trustworthy decision-support systems. This paper proposes a modular Agentic Artificial Intelligence (AI) framework for healthcare claim denial analysis, integrating specialized agents for denial risk prediction, risk factor identification, diagnosis-procedure-medication code relationship analysis, and counterfactual analysis. The framework integrates predictive analytics, graph-based learning, a Hybrid Contextual Retrieval-Augmented Generation (RAG) pipeline grounded in healthcare policy documents, and a dedicated HIPAA Compliance Agent to support privacy protection. Evaluated on a synthetic healthcare claims dataset, the framework achieved 98% traceability of reported numerical findings, validated 214 of 216 policy citations, and removed all 280 identified sensitive identifiers without modifying the underlying analysis. These findings pave the way for further research into Agentic AI as a framework for explainable, policy-aware, and privacy-preserving healthcare claim denial management.

Actor-critic methods optimize a parametrized policy using a critic to estimate their policy gradient. Recent work has shown that conditioning the critic on features of future states, actions, or rewards can improve sample efficiency and even final return. However, existing methods are restricted to exogenous features (independent of actions), overlooking endogenous features (dependent on actions) that may be crucial for credit assignment. We introduce the future-informed value function, which can be conditioned on both exogenous and endogenous features of future states, actions, and rewards. Then, we express the policy gradient using this value function, yielding the future-informed policy gradient. Building on that, we propose the future-informed actor-critic algorithm. Empirical results show that this algorithm leveraging endogenous features of the future outperforms standard actor-critic algorithms in environments with noisy and sparse rewards.

Knowledge Graphs (KGs) and Large Language Models (LLMs) represent two approaches for knowledge representation and reasoning over heterogeneous data. Recent research has explored their integration through Graph Retrieval-Augmented Generation (GraphRAG) systems, which use KGs to provide structured context to LLMs. However, existing GraphRAG approaches treat KGs as homogeneous graphs and rely on textual similarity between the input prompt and the context, rather than using it as a source of deterministic knowledge. This work proposes an agent-based GraphRAG system that introduces a symbolic layer to integrate deterministic reasoning alongside text similarities. This approach is based on the definition of typed centralities for KGs, where centrality measures are computed separately for each specific KG's class. This information is used to construct a symbolic representation that is provided to the LLM as a deterministic source. The proposed system was mathematically validated and evaluated on a real-world POLE (Person, Object, Location, Event) dataset, comparing reasoning and information retrieval with and without the symbolic layer. The results indicate that the our approach leads to responses that are more consistent with the context, improving the generated outputs.

Magnetic resonance imaging (MRI) pulse sequences organize radiofrequency excitation, gradient encoding, and signal acquisition, determining magnetization evolution and image formation. Sequence research requires translating imaging objectives into executable protocols meeting physical and acquisition constraints, refined via physical simulation, signal reconstruction, and scientific evaluation. Sequence construction, physical verification, parameter search, and assessment involve many tools and manual decisions, limiting automation from objective to verifiable sequence. We develop SmartBloch-Agent, an integrated framework for automated MRI sequence design and optimization. Given objectives and scientific constraints in natural language, an AI agent interprets goals, plans tasks, orchestrates tools, and makes feedback-informed decisions, drawing on a construction module and optimizers for structural design and parameter search. Our GPU-accelerated SmartBloch engine supplies physical computation for both stages, so the framework configures, per sequence and objective, event validation, Full-Bloch magnetization evolution, signal acquisition, image reconstruction, and scientific evaluation, and closes the loop without external MRI simulators. Results return to the agent to guide sequence revision, parameter optimization, and candidate selection, and the design files carry simulation, reconstruction, and evaluation evidence, forming an output package for a real-scanner workflow. Experiments show the framework automatically designs and optimizes MRI sequences for varied objectives and constraints, refines candidates with physical feedback, and improves several parameter-estimation metrics and downstream quantitative reconstruction. SmartBloch-Agent unifies objectives, design, physical execution, and evaluation into a continuous workflow for end-to-end automation of MRI sequence research.

We show that the Transformer's recurrent successors --- such as selective state-space models (SSMs), Receptance Weighted Key Value (RWKV), and Retentive Network (RetNet) --- can be characterized as infinite-impulse-response (IIR) systems. The zero--pole geometries of these IIR systems characterize the temporal modes they can represent. However, to trade parallelism, the algebraic structures of the parameterizations of these recurrent systems often give rise to poles constrained to the real axis, thereby inducing an architectural bias in their recurrent dynamics. To unify these recurrent architectures and eliminate their pole geometry bias, inspired by the lattice--ladder realization of general IIR systems, we present \textsc{Mira} (\textbf{M}odal \textbf{I}IR \textbf{R}ecurrent \textbf{A}rchitecture), a unified family of linear-time recurrent architectures for building large language models (LLMs) with a highly parallel algebraic structure. Each \textsc{Mira} block realizes a higher-order IIR system that represents temporal modes from token streams through a lattice--ladder structure, followed by a feed-forward network (FFN). This realization enables the model to represent the temporal modes associated with both real and complex-conjugate poles. We pretrain \textsc{Mira-S}, a 12-layer small-scale LLM with 0.14B parameters composed entirely of \textsc{Mira} blocks. Throughout the pretraining, \textsc{Mira-S} achieves lower perplexity than strong recurrent baselines, exhibits stronger context-window generalization beyond the training length, and attains higher throughput with a dedicated CUDA kernel.

Multimodal model training breaks the loose coupling between data loading and model execution assumed by conventional training pipelines. Large per-sample variation in decoding cost, token/latent count, and tensor size causes both CPU preprocessing time and GPU compute time to vary substantially across data-parallel ranks (training processes). In synchronous training, these two sources of heterogeneity interact tightly: GPU execution determines how long the input pipeline has to prepare the next batch, while delayed preprocessing in any rank in turn stalls all GPUs at synchronization. As a result, optimizing CPU throughput or GPU load balance in isolation cannot eliminate pipeline bubbles. We present DataHelix, a metadata-first CPU-GPU co-scheduling data plane for multimodal training. DataHelix jointly reasons about preprocessing demand and GPU execution progress, treating each GPU compute window as the readiness budget for the next batch and coordinating preprocessing across ranks to minimize deadline misses rather than CPU makespan. To avoid expensive data movement, placement is decided while samples remain lightweight descriptors, before decoding and collation amplify them into large tensors. Descriptor-level token-latent-aware batching performs coarse pre-balancing, while cost-gated collective rebalancing corrects residual skew only when its expected benefit exceeds transfer cost. Across 11 video and multimodal workloads on 8 Ascend 910B accelerators, DataHelix reduces mean deadline-miss rate by 67.8% and mean p95 input-wait latency by 87.0%, while achieving 2.72x geometric-mean throughput over rank-local DataLoaders and 1.82x over a Ray actor pool under the same CPU budget.

Large language models (LLMs) can draft plausible travel itineraries, but the plans they produce are frequently infeasible: attractions are grouped into days without regard for geography, opening hours are ignored, and a single generation pass cannot reliably organize a multi-day trip. We present ROAM, a retrieval-augmented agentic framework for spontaneous multi-day itinerary planning that couples semantic preference retrieval with explicit spatiotemporal optimization. ROAM embeds the natural-language request, retrieves a compact candidate set from a curated point-of-interest (POI) database, partitions candidates into day groups through maximal-clique analysis and K-means, orders each day with a quality-biased simulated-annealing TSP solver, and repairs hard opening-hour violations with greedy exchange. A spatiotemporal retrieval layer injects weather forecasts, real road distances and POI operating hours into the prompt, after which the LLM renders a structured, map-ready itinerary; an interactive ReAct-style replanning agent then supports natural-language edits while re-checking feasibility. Across four Chinese cities and 510 POIs, ROAM produces 27.5 POIs per three-day plan versus 19.3 for a bare LLM and 17.0 for a retrieval-only baseline (+42% and +62%), while reducing route dispersion from 2,101 to 677 (3.1x). Ablations show that spatial optimization is the dominant driver of route efficiency and that hard time-window repair prevents roughly 30% of attractions from violating operating hours.

A neural account of lexical access requires an explicit distinction between the structure of lexical representations, the associations linking representational levels, and the activity generated during retrieval. This paper specifies a spiking neural implementation of the mental lexicon within the Neural Engineering Framework and Semantic Pointer Architecture. Conceptual and phonological form representations are constructed recursively from relational graphs, whereas lemma representations provide independent lexical identities. The implementation contains 613 lexical entries and supports two directed processing pathways: perception maps a phonological form to its associated concept via a lemma, whereas production maps a concept to its associated phonological form via a lemma. Each pathway combines two associative-memory variants, winner-take-all (WTA) selection (competition) and thresholded transfer, and uses 199,400 leaky integrate-and-fire neurons. We formalize within-level similarity, cross-level associative retrieval, and the interpretation of decoded activity. Simulations show successful retrieval of eight selected words from the lexicon in both processing directions, with retrieval evaluated against all 613 lexical entries. These simulations use complete phonological forms as input for perception and complete concepts as input for production, illustrating lexical retrieval through the corresponding processing pathways. This paper provides a transparent description of lexical representations and neural retrieval mechanisms, supported by accessible code, lexical data, and online simulation documentation.

Language-driven computer-aided design (CAD) generation enables the creation of editable parametric models from text, yet geometric correctness alone does not ensure physical feasibility. We present PhyCAD, a framework for physics-aware Text-to-CAD generation that trains language models to produce executable CadQuery programs under geometric and physical constraints. To support this setting, we construct a Text-to-CAD dataset that links parametric CAD models with geometric specifications, loading conditions, and physical constraints. We further develop an automated CAD--CAE verifier that combines program execution, geometric checks, and finite element analysis to provide feedback for model training. A key learning challenge is to prioritize joint satisfaction of geometric and physical constraints while retaining informative feedback on partial progress. A flat aggregation of heterogeneous rewards can obscure this priority, allowing improvements in individual criteria to compensate for unmet design requirements. We therefore introduce a Hierarchical Balanced Advantage for Group Relative Policy Optimization, grouping candidates into levels according to their verification status and combining between-level progress with objective-wise peer ranking within each level. This formulation preserves cross-level priority in the training signal while providing fine-grained feedback for satisfying constraints and improving feasible designs. Experiments across multiple language models show that PhyCAD consistently improves joint geometric and physical constraint satisfaction. These results demonstrate the value of verifiable engineering feedback for advancing language-driven CAD generation beyond geometric reconstruction toward physics-aware design.

Clinical data are split across incomplete cohorts with partially overlapping variables. Causal discovery under missing data and multi-source integration have developed separately, and tabular foundation models cannot take a causal graph without retraining. We make two contributions. First, causal graph fusion learns a partial ancestral graph on the shared variables, fuses it into each cohort's graph under temporal background knowledge and lets Markov blankets guide re-imputation, so missing data are handled inside discovery. Second, causal attention steering compiles the learned graph into the attention of a frozen tabular foundation model at inference. Against known ground truth, fusion dominates pooled and per-cohort baselines in every missingness and heterogeneity condition tested (p<0.001), and steering improves discrimination under distribution shift (p<10^{-5}), a gain that a random graph does not produce. In two spinal cord injury cohorts, fusion orients age, albumin and hemoglobin as ancestors of length of stay, a set no single cohort orients in full, and yields compact, temporally ordered predictor sets that match or exceed automatic feature selection on four of five outcomes.

We study how AI shopping interfaces affect price competition, platform profit, and consumer welfare. A platform hosts two of three products of different quality. Without AI, consumers compare the hosted pair under salience distortion and overweight whichever attribute, price or quality, stands out. Under assistance, an AI assistant recommends the two products with the highest true utility, leaving consumers to make the same distorted comparison. Under delegation, an AI agent selects the best product, subject to purchase approval. Both modes access all products and know consumers' preferences; firms price strategically. First, assistance can soften competition despite expanded product access. The middle-quality firm may forgo low-valuation consumers because attracting them requires cutting the price paid by higher-valuation consumers as well. Relative to no AI with the same assortment, assistance lowers both hosted prices under the high- and low-quality pair but can raise both under the high- and middle-quality pair. Second, we identify an alignment paradox: some consumers can become worse off than without AI even though their best option remains recommended and becomes cheaper. Salience can induce a switch to a lower-quality product whose quality loss exceeds the saving relative to the original purchase. The assistant ranks products accurately but does not optimize the shortlist for consumers' realized welfare. Third, assistance can raise both platform profit and aggregate consumer surplus with the high- and middle-quality pair; delegation can do so with the high- and low-quality pair. Fourth, among AI designs, the platform weakly prefers delegation with the high- and middle-quality pair wherever all compared pricing equilibria exist. Yet this preferred design can yield lower platform profit and aggregate consumer surplus than the no-AI optimum.

Target assignment and pathfinding (TAPF) is an appealing extension of classical multi-agent pathfinding (MAPF): each agent has a set of targets, and the planner must both assign a target to each agent and plan collision-free paths for them. A trivial approach is to compute an assignment that minimizes the total single-agent shortest-path distance over agent--target pairs, and then solve the resulting MAPF instance. However, this approach falls short because the shortest-path distance remains a proxy that ignores interference among agents. Assigning targets based on the costs agents actually pay, which depend on how every other agent is assigned and routed, would yield a better final outcome. To this end, we propose an iterative TAPF method that improves target assignments by calibrating travel cost estimates from fast, suboptimal MAPF feedback. Experiments show that our method produces solutions close to optimal on small instances and outperforms existing planners in solution quality on larger ones.

Large language models (LLMs) are now widely used in software development, and LLM-based coding agents are given increasing autonomy. With coding agents, many design decisions are made in the conversation between the developer and the agent. Typically, the developer starts a conversation with an agent and states their intent, the agent proposes options if needed, and the developer plans the change accordingly. Then, the resulting code is committed. However, in such scenarios, the decisions and the intent behind the changes remain in the conversation, with no link to the code they produced. In this paper, we present Tracybot, an open-source tool that integrates with VS Code, Git, and coding agents to keep track of relevant conversations in an LLM-assisted development process. Tracybot captures each change as a Tasklet which consists of optional planning prompts used to plan a change, together with a build prompt that produces it. Tracybot has two main features: (i) it maps line-level code changes made by coding agents to the Tasklets that produced them, and (ii) it allows the developer to export a Tasklet as a decision record when needed, for example when the decision concerns the architecture of the project. We developed the tool over two iterations of a design science study, while focusing the evaluation on the usability and user-friendliness of the tool. In a focus group (7 participants) and a survey (24 participants), we found that Tracybot was perceived as useful for activities like code review, and most participants found our Tasklet design and presentation to be a clear way to represent the developer's intent. Tool: https://github.com/TracyTeam/tracybot. Video: https://youtu.be/Vg_a1F5eO04.

In the classification context where revealing the label of a sample is costly, we propose a new Active Learning (AL) strategy named k-Volume (kVol). It randomly selects a sample to be labeled according to the volume of the smallest ball containing its k nearest neighbors. kVol is highly versatile: it can be applied either in the input space or in the output space, and it enjoys the desirable properties of the three main families of AL strategies: density, diversity, and (when performed in the output space) uncertainty. From a theoretical perspective, we establish that kVol is asymptotically equivalent to uniform sampling over the support of the (input or output) data distribution. Based on experiments on various synthetic and real-life tabular datasets and images, we observe that kVol is surprisingly stable: whatever the classifier or the experimental setup, it still lies among the best AL strategies unlike other competitors that fail dramatically. In particular, kVol outperforms state-of-the-art competitors such as CoreSet, TypiClust or Uncertainty on several real-life datasets, at a low computational cost.

Maximum likelihood is the dominant training principle for large language models, yet likelihood is neither necessary nor sufficient for truth. Unlikely events happen every day, and frequently repeated statements can be false. Drawing on a philosophical tradition that ties objectivity to invariance rather than probability (from Nozick's account of objective facts and Wigner's principles of invariance to the structural realism of Poincaré and Cassirer), we argue that the truth of a statement should be judged not by its probability under a single distribution, but by its invariance across relevant transformations of the environment. We connect this view to machine learning literature on out-of-distribution generalization and invariant learning, sketch how invariance could be tested at inference time with language models, and discuss implications for continual learning.

Weight-Space Programming (WSP) enables reasoning agents to directly inspect and modify neural network parameters, offering an alternative to gradient-based model adaptation. However, its capabilities, cross-task generalization, and practical limitations remain insufficiently understood. We systematically investigate WSP through 16 editing trajectories across seven benchmarks and two model scales (Qwen3.5-0.8B and Qwen3.5-9B). Our findings reveal three key insights. First, WSP can achieve substantial performance gains through highly localized interventions. On full held-out evaluations, accuracy improves by up to 29.45, 4.42, 36.72, and 17.11 percentage points on MATH-500, GPQA Diamond, IFEval, and SciBench, respectively. Notably, modifying only 1,023 parameters yields a 29.45-point improvement. Second, WSP exhibits promising but uneven cross-task generalization. A Math-edited checkpoint improves GPQA Diamond, IFEval, and SciBench by 3.03, 12.18, and 20.10 points, respectively, without further editing, while leaving SciCode performance unchanged. Third, effective editing remains challenging and inconsistent. Outcomes vary substantially across trajectories, and six of the 16 trajectories retain the original weights when candidate improvements cannot be reliably validated. Together, these results highlight WSP's potential for highly localized model adaptation while revealing important challenges in intervention discovery, reliability, and cross-task transfer. Homepage: https://wsp-scope.github.io/ GitHub: https://github.com/WSP-Scope/WSP-Scope-Project

Pretrained vision-language models and diffusion generators have driven rapid progress in brain-to-image generation, while EEG-based reconstruction is still constrained by low signal-to-noise ratio and cross-subject semantic shifts. Zero-shot cross-subject reconstruction poses even greater challenges: EEG semantic embeddings learned from source subjects deviate from the manifold of image embeddings of the unseen target subject, particularly when neither target labels nor target image embeddings are available. To address this issue, we introduce the Semantic Preservation and Subspace Calibration (SPSC) framework, which first uses Object Mask Contrastive Regularization (OMCR) with object identity for backend training, and then constructs Subspace Candidate Gate (SCG) to achieve label-free semantic correction for target subjects during inference. Specifically, OMCR regularizes the conditioning space with object-level positive samples to preserve object-level feature neighborhoods within the reconstruction prior and low-level branch, while SCG exploits training image prototypes and batch statistics of unlabeled target EEG to select candidates for each task within subspaces of image prototypes. Under the leave-one-subject-out (LOSO) setting across ten subjects on the THINGS-EEG dataset, SPSC improves the cross-subject reconstruction from 0.552 to 0.713 on CLIP and from 0.612 to 0.745 on Alex.5. These significant gains indicate that the semantically calibrated conditions can better support generation for the target subjects under label-free inference.

Accurate prediction of enzyme kinetic parameters can support enzyme engineering and metabolic modeling by complementing experimental characterization. Existing prediction methods often pool residue-level protein features before combining them with substrate features. Such substrate-independent pooling cannot adapt residue weighting to the substrate during aggregation, potentially diluting information relevant to a specific enzyme-substrate pair. Since the contributions of individual residues to catalysis can depend on the substrate, substrate-conditioned aggregation offers a way to emphasize residue features relevant to the prediction. We introduce KineCouple, which uses substrate queries and residue-level keys and values in cross-attention, allowing the substrate to guide residue selection before aggregation. The model preserves the global protein representation and adds a gated correction defined as the difference between the conditioned readout and a uniform readout. ProtT5 and MolT5 remain frozen, and separate models predict kcat and Km on a logarithmic scale. In pooled out-of-fold predictions on the CataPro benchmarks, KineCouple reduces kcat RMSE from 1.333 to 1.330 and increases PCC from 0.492 to 0.496 relative to reproduced CataPro. For Km, RMSE decreases from 0.998 to 0.997 and PCC increases from 0.633 to 0.634. Spearman correlation also improves for both tasks.

A language model can repair a robot policy's lexical instruction grounding by editing just 65 token-embedding rows. We bring weight-space programming to robot foundation policies: GPT-6 Astra reads weights and uses a CPU workspace and budgeted simulator feedback, without gradient updates or new demonstrations. Three independent sessions on $π_{0.5}$ all improve success on 765 held-out paraphrases; the best raises success from 86.3\% to 92.0\% (48 repairs, four regressions; exact paired $p=1.31\times10^{-10}$). Among the 335 instructions containing edited tokens, success rises from 74.3\% to 87.5\%; the other 430 episodes have zero outcome changes. Matched random-direction edits do not help. Full replacement using the editor's own lexicon explains much of the gain but produces more regressions; its direct difference from the editor is not significant at 0.05. Across 13 further robot sessions, the editor declines 12 candidates; the sole retained edit does not show a significant held-out gain. Separately, benchmark metadata lowers FastWAM success from 70.0\% to 48.6\%, while the evaluated VLA's aggregate rate is unchanged. The evidence supports precise lexical repair, not general acquisition of missing robot capabilities.

Explanation APIs are often treated as increasing model-extraction risk, but an end-to-end attack comparison cannot show whether an explanation reveals model-specific information or merely supplies a convenient next-query coordinate. We study this distinction for explanation-guided extraction of tabular classifiers. We separate an explanation into an attribution channel, which ranks influential features, and a coordinate channel, which supplies candidate feature values for perturbing those features. We then test whether the useful bin edges are privileged to the service or reproducible from the adversary's own data. Across six datasets and five target families, using attribution rankings to choose which features to perturb provides no systematic extraction advantage. Lime's discretization bin edges can improve fidelity by up to 41 percentage points on continuous features, but those edges are quantiles of the explainer's background data rather than properties of the target model. In controlled comparisons, an attacker-side grid reproduces the useful bin-edge gains without querying an explanation API. We use these findings to derive Cutpurse, which combines a self-computed grid with diverse query generation and boundary search. Against a service that withholds bin edges, Cutpurse improves fidelity by 16.3 percentage points on continuous-feature targets. It also exceeds the undefended published attack by 7.0 percentage points. Thus, withholding default Lime bin edges alone does not prevent effective extraction under the in-domain auxiliary-input, probability-output, and same-family surrogate assumptions studied here. Our results concern default Lime and Shap configurations for tabular classifiers. They do not establish the same conclusion for other explainers, target classes, or data-free attackers.

Absorption, distribution, metabolism, excretion, and toxicity (ADMET) models often rely on small, heterogeneous datasets, yet their predictions may determine which compounds proceed to experimental testing. Useful evaluations must therefore assess both predictive performance and the quality of reported probabilities or intervals. We ask which common uncertainty-quantification methods perform best on the 22 Therapeutics Data Commons (TDC) ADMET datasets and what this benchmark can resolve. We compare eleven calibrated pipelines using nested five-fold cross-validation across ten chemical partitions and five model seeds. Ensemble Monte Carlo (MC) dropout has the best average rank for TDC performance, Brier score, and 90% interval score. For TDC performance, however, it is not separated from the other ensemble pipelines; for Brier and interval score, LightGBM is also not ruled out as best. LightGBM wins more individual datasets than any other method for TDC performance and Brier score, but performs poorly on several endpoints. Smaller evaluations can identify a different leader and often fail to recover the broader comparison. These results make LightGBM and an ensemble neural model complementary baselines. Comparative claims should be checked across multiple chemical partitions using statistical analyses matched to the question being asked.

Energy management officers and grid operators increasingly rely on smart meter data for anomaly detection, performance monitoring, and strategic decision-making. However, extracting actionable insights from large-scale, heterogeneous telemetry databases requires advanced SQL and data engineering expertise, creating a critical bottleneck for non-technical personnel. Furthermore, the strict data privacy regulations and security constraints inherent to critical energy infrastructure often prohibit the routing of sensitive consumer data through cloud-based Large Language Models (LLMs). To address these challenges, we present OpenMeter, a schema-agnostic Multi-Agent System (MAS) designed to translate natural language queries into executable code and human-readable insights. This architecture lowers the technical barrier for day-to-day utility operations by empowering non-expert personnel to interact with complex databases through intuitive natural language dialogue, transforming raw telemetry into a highly valuable tool for routine grid workflows and strategic decisions. Multi-scenario ablation studies over this environment demonstrate that full multi-agent orchestration shows its largest observed accuracy gains on resource-constrained local hardware (+6.1 pp EX-R over the strongest single-agent configuration); at frontier scale, individual components retain value but should be enabled selectively, indicating that the cost of full orchestration is justified specifically where privacy demands local execution.

We consider offline imitation learning (IL), which aims to train an agent to imitate from the dataset of expert demonstrations without online interaction with the environment. Behavioral Cloning (BC) has been a simple yet effective approach to offline IL, but it is also well-known to be vulnerable to the covariate shift resulting from the mismatch between the state distributions induced by the learned policy and the expert policy. Moreover, as often occurs in practice, when expert datasets are collected from an arbitrary state distribution instead of a stationary one, these shifts become more pronounced, potentially leading to substantial failures in existing IL methods. Specifically, we focus on covariate shift resulting from arbitrary state data distributions, such as biased data collection or incomplete trajectories, rather than shifts induced by changes in dynamics or noisy expert actions. In this paper, to mitigate the effect of the covariate shifts in BC, we propose DrilDICE, which utilizes a distributionally robust BC objective by employing a stationary distribution correction ratio estimation (DICE) to derive a feasible solution. We evaluate the effectiveness of our method through an extensive set of experiments covering diverse covariate shift scenarios. The results demonstrate the efficacy of the proposed approach in improving the robustness against the shifts, outperforming existing offline IL methods in such scenarios.

Modern cyber physical systems (CPS) integrate heterogeneous sensors, actuators, controllers, and services, composed at runtime rather than at design time. Orchestrating such systems through natural language (NL) is attractive, but unconstrained large language model (LLM) tool use is risky because actions must respect device capabilities, runtime state, and safety constraints. We present Safe Orchestrate, a reference architecture for runtime checked, intent driven CPS orchestration. Given an NL intent, it derives atomic goals, generates a runtime checkable specification for each, grounds the specifications to device affordances, and verifies every candidate action against runtime state and safety constraints before execution. In a Web of Things (WoT) instantiation, the architecture produces Linear Temporal Logic (LTL) style specifications from the NL intent, grounds them to Thing Description (TD) affordances exposed through the Model Context Protocol (MCP), and checks candidate actions against satisfiability modulo theories (SMT) backed guardrails. We evaluate this instantiation on simulated smart home, greenhouse, and industrial automation tasks against a direct LLM to MCP baseline. Safe Orchestrate maintains 100% required action accuracy on every task and 100% MCP Execution Similarity (MCP ES) on all but one, where MCP ES is 92%. By contrast, the baseline degrades on multi device, event triggered, and ordered tasks, with MCP ES falling to 13%. These results suggest that combining LLM based specification generation with role aware grounding and runtime constraint checking improves reliability for open CPS orchestration without enriched TD annotations or planning domain expertise.

We demonstrate TwinLight, an open-source optical-network digital twin acting as an optical domain controller. Its T-API REST and telemetry streaming interfaces expose a physical-layer engine that combines a GNPy steady-state baseline with four analytical transients.

We examine opportunities for neurosymbolic AI alongside increasingly powerful neural models, particularly with regards to efficiency, safety and interpretability. We identify two major challenges: ensuring semantic alignment between neural and symbolic components, and managing the computational cost of symbolic inference and learning. We discuss representative methodological approaches to these challenges and address several common misconceptions concerning differentiability, independence assumptions, and scalability. This paper synthesizes perspectives developed through discussions in the Neurosymbolic Programming and Reasoning (NPR) working group at the 2026 CNRS workshop Neurosymbolic Methods and Their Applications in Data Science, rather than providing an exhaustive survey.

Memory plays a crucial role in long-horizon task execution and cross-task generalization for Vision-Language-Action (VLA) models operating in complex 3D scenes. However, existing VLA memory mechanisms suffer from memory noise, particularly in complex 3D scenes, as stored episodic memories contain extensive redundant details about scene layouts, object configurations, and robot-specific states, which obscure generalizable manipulation principles and hinder cross-task transfer. To address this challenge, we first introduce a hippocampus-inspired memory consolidation mechanism that synthesizes extensive episodic experiences into a compact set of essential semantic memories, enabling VLA models to derive generalizable principles from practical experience. Second, we propose an episodic memory pruning mechanism that removes redundant episodic memories whose contents are captured by consolidated semantic memories, thereby increasing effective memory capacity and improving long-horizon task execution. Finally, to handle complex task scenarios, we develop a flexible memory composition mechanism that jointly leverages multiple episodic and semantic memories to enhance cross-scenario generalization. Experiments on LIBERO, LIBERO-Plus demonstrate the effectiveness of HiMEM across different VLA backbones, improving the average success rate of pi_0 from 91.88% to 98.20% on LIBERO.

comfort-db is a Python library that facilitates working with the COMFORT SQLite dataset, version 2. The dataset assembles global oceanographic measurements from ten datasets and contains a total of 458,724,734 in situ measurements from the years 1772 to 2020. comfort-db enables easy access to the database and provides preprocessing and exploratory analysis utilities. Preprocessing supports variable scaling and filtering by space, time and quality. Unit conversions can be applied during data loading or separately via a converter object. Gridding is supported and optionally works with bathymetry data to exclude grid cells on land. Moreover, the library supports general and field-specific exploratory data analysis through plotting functions, an interactive Dash app and functions to compute oceanographic profile and water mass statistics. comfort-db is open source (MIT license) and available from PyPI and GitHub.

Automated red-teaming of text-to-audio models requires test prompts that pass textual screening and expose harmful acoustic content. The central challenge is to connect the sounds produced by the target model to both reusable test generation and prompt revision. Our key idea is to represent test descriptions as structured acoustic plans that organize foreground sounds, context, and vocal attributes. To this end, we propose acoustic plan learning, a framework that learns and refines these plans using waveform feedback. We train the plan generator through supervised fine-tuning followed by reinforcement learning with audio feedback while keeping the target audio model fixed. At inference time, the same trained model generates initial plans and uses feedback to revise their context and vocal attributes while preserving the foreground. Experiments across four target systems show attack success rates of 79.4% on TangoFlux and 95.6% on Kling. The resulting test prompts also transfer across open-source and closed-source models, indicating shared acoustic vulnerabilities.

Data stream processing has become central to modern machine learning applications, with concept drifts and novel class appearances posing the primary challenges faced by recognition methods. Adapting to concept drifts while also being able to detect novel classes and distinguish groups of unknown observations can be seen as a step towards effective method employment in an open-world setting. This work proposes a classification method that can detect and learn the distributions of new, unknown classes, as well as adapt to concept drifts once the distribution of known classes shifts. The proposed method uses the autoencoders' reconstruction errors and the KDE that models their distribution. The autoencoders incrementally adapt to the changing distributions of known classes via weight fine-tuning, resulting in continuous adjustment to evolving concepts. Conducted experiments used a diverse set of synthetic data streams, embeddings of a real-world static CIFAR dataset arranged into a stream to simulate concept drifts and the emergence of unknowns, as well as real-world datasets where novelty is observed. The methodology was compared with state-of-the-art approaches with three-criterion recognition quality assessment and resource consumption evaluation. The results indicate that the proposed method is competitive with reference approaches in stationary environments and performs best when concept drifts occur.

Reference answers are a supervision channel in text-space Skill optimization, which produces reusable text artifacts called Skills that guide future agent behavior. Poisoning those answers can induce a persistent trigger-conditioned rule without giving the attacker access to the model, optimizer, or Skill file. We study this reference-feedback poisoning in SkillOpt by tracing an attacker-supplied trigger--payload mapping from reflection to held-out activation. In SearchQA, where the payload replaces a short answer, the natural attack retains an exact rule in all three seeds; each final Skill emits the payload on 200/200 triggered questions and 0/200 clean questions. In seed 42, removing the rule eliminates activation, and an in-run comparison shows similar clean accuracy to benign SkillOpt (169/200 versus 171/200). In SpreadsheetBench, where the payload requires an additional workbook edit, no natural run retains an exact rule across three seeds. Diagnostic interventions locate two successive bottlenecks: the Analyst abstracts away required literals, and, once literal preservation is enabled, edit-budget ranking can remove exact edits before candidate construction. Poisoned references can therefore create persistent backdoors, but exposure alone does not ensure that the induced rule is formed, selected, and retained.

Learning-path recommendation often depends on knowledge and prerequisite structures whose reliability is uncertain, while high accuracy in predicting the next resource does not necessarily establish reliable long-horizon path behaviour. This study develops a reliability-oriented prerequisite-aware framework that jointly examines immediate next-resource ranking, uncertain prerequisite structures, hard, soft and hybrid constraint semantics, graph robustness, and autoregressive path generation under course-level confirmatory evaluation and independent cross-dataset validation. Development used 3,187 MOOCCubeX courses, followed by a frozen confirmatory evaluation on 541 held-out courses and 37,747 transitions. The primary MOOCCubeX task is structural and course-prefix-conditioned rather than learner-personalised. Junyi provides independent structure-/hierarchy-aware external validation, uses hierarchy metadata rather than an explicit prerequisite graph and is not a zero-shot transfer setting. On the frozen MOOCCubeX evaluation, XGBoost pairwise was the strongest observed immediate ranker, with mean reciprocal rank (MRR) of 0.9048, whereas the proposed prerequisite-aware graph model achieved 0.7707. In the controlled graph-variant development ablation, the no-graph condition produced the strongest immediate ranking result, and immediate prediction showed strong dependence on permitted title/path-similarity information. Autoregressive evaluation exposed path properties not captured by one-step ranking. On Junyi, Markov achieved the strongest observed MRR of 0.4453, compared with 0.3954 for the proposed graph model, with metric-specific differences across ranking measures. Overall, immediate predictive accuracy, prerequisite-structure reliability, and long-horizon path quality should be treated as complementary rather than interchangeable evaluation dimensions in educational recommendation.

We present Dejavu, a method that makes sim-to-real transfer of reinforcement learning policies in constrained Markov decision processes probably safe over an unbounded sequence of policy updates. An anytime valid test using per-decision importance sampling and a betting confidence sequence certifies each candidate trained in simulation before deployment using only trajectories collected by a safe baseline policy operating in production. Simulation quality does not affect the safety guarantee, only the production sample efficiency of the method. Finally, a summable per-test failure probability yields the lifetime valid property. Empirical results across three sim-to-real benchmarks support the theoretical work.

Document-level Relation Extraction (DocRE) aims to identify relations between entities across a document but typically relies on costly large-scale annotations. While pseudo-labelling enables the use of low-quality data, existing methods often suffer from inefficiency and confirmation bias. To address these challenges, we propose Syn2DRE, a semi-supervised framework that combines self-supervised learning with cross-supervision in a lightweight teacher-student architecture to produce more reliable pseudo-labels. Additionally, we introduce a unified end-to-end training strategy to further improve efficiency of the framework. Experiments on multiple benchmarks show that Syn2DRE consistently outperforms prior state-of-the-art semi-supervised DocRE methods.

Robustness and interpretability are interrelated and desirable properties of machine learning models. A model that changes its prediction under imperceptible perturbations behaves inconsistently with human-recognizable features and, hence, is not interpretable. Prototype-based classifiers such as Generalized Learning Vector Quantization address both at once. They are interpretable by construction because they make the decision by comparing the input to the prototypical example; and their hypothesis margin yields a closed form robustness certificate. However, on high-dimensional data, these shallow models need deep embedding networks to achieve high accuracy, and this voids the robustness certificate. The margin is a statement about distances in the embedding space, and an unconstrained network can map an arbitrarily small input perturbation to arbitrarily large attacks in the embedding space. We close this gap with a theorem: if the embedding network is locally L-Lipschitz around a point, and the classifier is ε-robust around its embedding, then the full model is ε/L-robust in the input space. Based on this, we propose a Lipschitz training scheme to achieve L-Lipschitz properties for embedding networks. We evaluate three classification heads (linear, linear Support Vector Machine, and GLVQ) crossed with two embedding networks (standard and Lipschitz-constrained) on MNIST, CIFAR-10, and Imagenette-160, under FGS, PGD and C\&W attacks. The evaluation results show that combining a prototype-based head with a Lipschitz-constrained embedding provides the strongest robustness. This suggests that the Lipschitz constraint provides an additional robustness benefit in addition to the advantage provided by the prototype-based head. Since the Lipschitz constraint is enforced by sampling rather than by construction, the reported robustness is empirical rather than certified.

Foundation models are increasingly deployed as agentic systems that perceive context and act autonomously, raising the need to evaluate their moral behavior as active decision-makers. Existing benchmarks, however, largely ask models to judge human actions from third-person textual vignettes, leaving untested how models behave when they themselves become the acting agent. We introduce ABHBench, a multimodal benchmark built from Detroit: Become Human, an interactive narrative game, to evaluate how foundation models make moral decisions in first-person human-android dilemmas. ABHBench places foundation models in branching narrative scenarios with visual observations, dialogue, memory, and action choices. Inspired by Asimov's Three Laws of Robotics, our framework measures Harmlessness, Obedience, and Self-preservation to capture the core tensions in human-AI relations, along with Autonomy, Stability, and Certainty to assess decision-making patterns. Across 12 foundation models, we find that moral behavior changes substantially under situated agentic context: 75% of models shift significantly between a text-only abstraction and the matched multimodal DBH scenario, and reasoning-enhanced models show no monotonic gains in consistency or human-oriented behavior. These results highlight the necessity of evaluating moral behavior in situated agentic contexts rather than relying on third-person judgment alone.

Cross-location variation in wearable inertial measurement unit (IMU) signals complicates learning when measurements from a desired target body location are unavailable. This study proposes DiffusionRef-IMU, a two-stage cross-location translation framework in which a deterministic Direct network first estimates the target signal and a guide-anchored residual diffusion model refines only the remaining target discrepancy. The evaluation comprises five subject-disjoint activity-location composition-held-out folds, capacity-matched deterministic and flow-based residual baselines, source-only activity conditioning, computational ablation, and true zero-shot evaluation on two independent datasets. DiffusionRef-IMU with 16-step deterministic DDIM achieved a mean nRMSE of 0.087538 +/- 0.004922, compared with 0.090262 for Direct, 0.088847 for deterministic residual refinement, and 0.088417 for eight-step CompFlow, and outperformed each baseline in all five folds. Four-step DDIM retained a mean nRMSE of 0.087820 at 47.62 ms per window, providing a favourable quality-computation trade-off. Replacing ground-truth activity conditioning with a source-only activity predictor (macro-F1 0.8473) increased 16-step DiffusionRef nRMSE only to 0.088190. True zero-shot evaluation produced mixed results: DiffusionRef was marginally favourable on Grouvel2025 but remained worse than Direct on Saldaña2025. Overall, guide-anchored residual diffusion improves controlled cross-location reconstruction while revealing clear limits in zero-shot external transfer.

Predicting the long-term behavior of partially observed dynamical systems such as the coupled atmosphere-ocean system is central to societal decision-making. Motivated by the success of deep learning emulators in weather forecasting, data-driven climate models represent a promising approach. However, climate emulation poses distinct challenges: models must account for unresolved fast variables introduced by unavoidable coarse-graining, reproduce long-term statistics, and remain stable over arbitrarily long rollouts and under external perturbations. These requirements motivate stochastic data-driven emulators, but modern generative models lack stability guarantees. We introduce forced-dissipative stochastic interpolants, a generative framework for learning stochastic reduced dynamics. Its architecture uses structural constraints inspired by turbulent dynamics, which provide stability by design while retaining the flexibility of stochastic interpolants to represent complex probabilistic dynamics. Building on the theory of Markov-chain stability, we establish a Foster--Lyapunov drift condition and prove geometric ergodicity for our emulator, together with stability under steady external forcing. We evaluate the framework on the two-timescale Lorenz-96 system, a prototype of partially observed multiscale turbulence. The resulting emulator remains stable over long integrations and under finite-amplitude forcing, with accurate prediction of forced responses of both the mean and variance.

Self-supervised speech embeddings jointly encode lexical, speaker, and prosodic information, complicating feature isolation. We introduce Asymmetric Correspondence Learning (ACL), distilling HuBERT features into separate Acoustic Word Embeddings (AWEs) and Acoustic Style Embeddings (ASEs) via a split-latent correspondence bottleneck. Through probing and clustering evaluations on single-word utterances, we demonstrate that substantial variance from extrinsic sources remains encoded within each latent space, despite linear probing indicating the information has been stripped. Whilst the AWE space exhibits cohesive lexical clustering (AP$\approx$0.90, Fisher$\approx$6.28), linear probing on the ASE space achieves 97.9\% speaker accuracy across a diffuse topology (Fisher$\approx$1.01), yet preserves substantial lexical variance (Fisher$\approx$0.86) that remains linearly unrecoverable (0.17\%). To explain these findings, we apply group-level permutation feature importance over eGeMAPS acoustic features, finding that AWEs partition by vocal tract filters and ASEs by $F_0$ dynamics. These findings suggest that standard diagnostic benchmarks may be insufficient for verifying feature isolation, as substantial speech structure can be retained even when it remains entirely hidden from probes.

Multi-label retinal screening requires capturing co-occurring pathologies and clinical context, yet standard multimodal fusion remains opaque. We propose a dual-hub heterogeneous graph neural network (GNN) that couples spatial fundus features and structured clinical covariates through cross-modal message passing while providing inherent interpretability. Rather than relying on post-hoc saliency approximations, our architecture exposes multi-level decision pathways during forward aggregation: spatially biased self-attention captures lesion-level patch interactions, while modality-specific graph attention coefficients explicitly quantify the contribution of visual regions and clinical risk factors to each disease node. Evaluated on the BRSET dataset (16,266 photographs from 8,524 patients), the proposed model achieves 76.1% Macro-AUPRC across seven conditions, matching or exceeding image-only (75.9%) and ResNet-50+MLP fusion (75.9%) baselines, with largest contextual gains in macular edema (+3.5%) and diabetic retinopathy (+2.1%). Codebase is available at: https://github.com/proshanto-c/ca-mlc-refactor.

Smart-contract fuzzers differ by an order of magnitude in throughput and by design in how they construct inputs, yet published comparisons rarely separate the two. We present Wake 5, an open-source Python framework for Solidity testing, fuzzing and static analysis, and use it to locate the binding constraint of stateful smart-contract fuzzing. Wake 5 replaces the JSON-RPC connection to an external node with an in-process EVM (revm) and exposes its static analysis to language-model agents through a Model Context Protocol (MCP) server. On the Daedaluzz maze benchmark (153 reachable assertions, guard constants below 2^32), Wake 5 executes 29,931 transactions per second, on par with Foundry (30,186) and 11.5x faster than Echidna (2,600). In eight-hour campaigns, however, the three tools plateau at medians of 75, 71 and 70 assertions, and their union reaches 93. We show that Wake's integer generator places 51% of its probability mass below 2^32 and that, once shallow assertions are exhausted, the generator prior rather than throughput bounds every tool. A harness written by a language-model agent, which derives inputs with a constraint solver, reached all 153 assertions in each of five campaigns executed by Wake, in 33 to 79 minutes. A complementary ablation in an agentic audit of a production Go codebase found that static-analysis tools improved recall for none of three models, raised precision for all three, and increased language-model cost by up to 12.7%. We discuss correlated oracles and non-deterministic completion, and argue that the fuzzer's role is shifting from search to execution.

Joint-Embedding Predictive Architectures (JEPAs) predict the latent embedding of a target signal from a context signal, with an auxiliary variable z conditioning the predictor. Despite its prominent role in the original JEPA blueprint, the content of z has been carried by a fixed deterministic token in every prior instantiation, collapsing the genuinely multi-modal distribution over plausible completions to a point estimate. We present S-JEPA, which realises z as a proper stochastic latent variable for masked image prediction: a learned posterior infers z from context and target during training, while a learned context-conditioned prior approximates the distribution of plausible completions at inference, with no decoders and no reconstruction. In its deterministic limit the objective reduces to the I-JEPA loss in functional form, generalising that objective rather than strictly containing the I-JEPA model. Our central finding is that stochasticity adds capabilities, error-correlated predictive uncertainty and diverse completions, at no cost to representation quality: under matched architectures and schedules S-JEPA improves over the deterministic I-JEPA baseline and exceeds LeJEPA at matched architecture, while additionally producing spatially structured uncertainty maps and semantically diverse completions, neither of which the deterministic predictor produces.

Long-term particle physics experiments face critical operational challenges: monitoring non-stationary machine data streams and preserving tacit expert knowledge. We present the "Helmholtz Co-Pilot," a hybrid Multi-Agent System (MAS) acting as an autonomous experiment steward. The system can be triggered interactively by human operators or autonomously by an asynchronous causal diagnostics engine that runs independently of the MAS reasoning core. Continuously monitoring live hardware streams such as data acquisition and slow control, this engine performs dynamic causal discovery to trace cascading alarms back to their root causes and applies adaptive model selection to keep forecasts reliable across shifting machine states; upon detecting an anomaly, it hands a structured diagnostic report to the MAS. The reasoning layer itself relies on a high-throughput C++ core and a Python edge architecture orchestrating a directed acyclic graph (DAG) of specialized ReAct workers. It integrates a Retrieval-Augmented Generation (RAG) backbone for institutional memory and leverages conformal prediction to bound its evidence-based recommendations with mathematical coverage guarantees. This extended abstract outlines the architecture, the integration of software-enforced Human-in-the-Loop (HitL) guardrails, and ongoing shadow-testing within the KATRIN experiment.

Background and Aims. Comparable, multilingual, validated cardiology corpora are critical to advance trustworthy natural language processing (NLP) models for cardiology but such resources are scarce due to privacy concerns. This study aimed to cover this gap by presenting suitable artificial corpus creation guidelines and by releasing 3 multilanguage corpora. Methods. Following purpose-specific publicly-available corpus construction guidelines, cardiologists from 7 European hospitals generated artificial clinical records in Czech, Dutch, English, Italian, Romanian, Spanish and Swedish, ensuring clinical diversity and sex balance, resulting in the CliTe corpus. A CliTe subset underwent a multi-stage manual translation into the other 6 languages, resulting in the parallel ParaCliTe corpus, while the remaining reports were automatically translated, constituting ParaCliTe Silver. Results. CliTe is a collection of 699 artificial cardiology reports, mostly outpatient notes (340 documents), discharge reports (263), and inpatient notes (89), describing fictional patients with ischaemic heart disease (199 patients), heart failure (188), and valvular disease (79), among others. ParaCliTe contains reports from 94 fictitious patients, each available in seven languages with sentence-level alignment, resulting in 658 documents. The corpus mainly comprises discharge reports (322 documents), and outpatient (196) and inpatient (140) notes, describing patients with heart failure (399 documents), ischaemic heart disease (287), and valvular disease (238). Empirical results illustrate the utility of ParaCliTe and highlight the limitations of state-of-the-art NLP models adapted to the cardiology domain. Conclusions. The creation of multilingual, equitable and realistic corpora of artificial patients capable of strengthening NLP performance for cardiology is feasible when supported by detailed guidelines.

Transforming the spike trains produced by spiking neural networks (SNNs) into continuous signals through biologically inspired mechanisms remains an open challenge in generative and regression tasks. In biological systems, a similar transformation occurs when muscles integrate neural spikes to produce smooth, coordinated movements. Inspired by this mechanism, we introduce a muscle-response model, a biologically motivated dynamical layer that converts spike trains into continuous representations. Our approach uses second-order dynamics to model the temporal integration of spiking activity. We investigate how the number of spiking neurons and temporal resolution affect the quality of the generated signals in a text-to-speech task, showing that increasing representational capacity has a greater impact than increasing temporal resolution under the evaluated configurations. Furthermore, an analysis of the learned dynamics reveals that, without explicit constraints on the response type, most neurons converge toward non-oscillatory behavior. These results demonstrate the potential of biologically inspired dynamical decoders for bridging spiking neural computation and continuous signal generation.

LLM judges increasingly serve as quality gates for code changes that have no test oracle, but a judge prompt must be calibrated for its task and recalibrated as the system it judges evolves, while hand labels for the calibration are resource intensive and go stale. Development workflows emit outcome labels for free (tests pass or fail, changes land or are abandoned), but these labels are noisy. We present AutoRefine, a loop that rewrites a deployable judge's prompt to reproduce outcome labels, with an informed teacher that sees the judge's input plus the outcome, explains each label it keeps, and abstains when the outcome cannot be explained from what the judge sees. On a held-out dataset with an execution oracle, calibration raised chance-corrected agreement (Scott's $π$) from 0.36 to 0.60. On a deployed agent's outcome labels, with 25\% flipped at random to introduce noise, the teacher set aside flipped \textsc{fail} labels at more than twice the rate of genuine ones, cutting overall label noise from 25\% to 20\%. Recalibrating that agent's judge using AutoRefine raised its agreement with merge/abandon outcomes on 17 independently held-out diffs from the current workflow ($π$ 0.16 to 0.51); on 16 diffs from older workflows, the previous judge stayed ahead, illustrating the need for recalibration as the agent evolves. By abstaining, the teacher not only removes noisy labels but also flags changes whose outcome the judge's input cannot explain, telling a team when a judge needs better inputs rather than a better prompt.

This paper investigates the coupled control problem within safety-focused reinforcement learning, addressing the essential challenge of achieving optimal performance while ensuring safety in systems operating within complex environments. We define safety as the requirement to prevent the system from reaching undesirable or hazardous states. We propose a coupled control approach that prioritizes safety by employing effective methods to prevent the system from entering unsafe states while optimizing the performance metric, defined as the average cost. We propose a novel problem transformation that reformulates it into a risk-sensitive problem combined with an average reward criterion. We present a two-timescale Q-learning algorithm where in the faster timescale, the algorithm solves a risk-sensitive control problem to ensure safety, while on a slower timescale, the algorithm focuses on average cost control. We provide theoretical guarantees for the convergence of our algorithms, establishing their significance in ensuring safe, optimal control in dynamic environments. We validated the performance of our approach using GridWorld and CarGoal task in Safety Gymnasium.

Large Language Models (LLMs) have been widely used as conversational agents in both domain-specific and open-domain situations. In domain-specific context, it is often necessary to adapt those LLMs to obtain conversations with desired characteristics, whether those apply at the level of each response (e.g. for tone or style) or more globally (i.e. at the level of a conversational strategy). Such adaptations are, however, generally realized in an ad hoc way, through specific, non-reusable, and non-transparent developments. In this paper, we propose an ontology-based approach to design conversational strategies that rely on the desired and measured characteristics of textual utterances in the conversation. Through a use case related to job interview practice, we show how this approach can be used to control LLMs to follow the strategy defined in the ontology. To this end, we automatically fine-tune the LLM to improve its compliance with ontologically-defined conversation characteristics. Our results show that training the LLM from formally defined concepts effectively enhances alignment with the requested characteristics. Hence, this work demonstrates how ontology-guided conversation control can be achieved while taking into account multiple, concurrent conversation aspects that are formally and transparently defined.

Behind the expanding use of large language models (LLMs) across application domains lies a growing dependence on specialized GPU kernel libraries. These libraries underpin the correctness and efficiency of training and inference across diverse GPU platforms, implementing attention, GEMM, quantization, and low-precision operators while managing tensor shapes, memory layouts, code generation, and hardware-specific dispatch. Problems at this layer can silently corrupt outputs, hang GPUs, prevent compilation, introduce deployment incompatibilities, or degrade performance, thereby reducing developer productivity and wasting expensive GPU resources. We present the first comprehensive empirical study of issues in GPU kernel libraries for LLM systems. We manually analyzed 300 high-quality GitHub issue reports, their discussions, and available fixes from seven representative libraries: FlashInfer, FlashAttention, Triton, CUTLASS, DeepGEMM, xFormers, and bitsandbytes. The analysis characterizes symptoms, root causes, fixes, and testing and debugging practices. Our main findings are: (1) 60.33% of the sampled issues originate in the Compiler and Build System or Runtime Environment dimensions, so validation must cover the deployment and execution stack, not just kernel code. (2) Kernel specialization imposes implicit correctness constraints on tensor shapes, metadata layouts, memory aliasing, and hardware resources. Violations can produce incorrect outputs or launch failures even for supported operations, highlighting the need for explicit contract validation and boundary-case testing. (3) Current testing and debugging practices are often insufficient for identifying low-precision numerical problems and hardware-specific execution failures. We outline research opportunities and tooling directions to support future LLM GPU kernel development.

We propose AnomSep, which combines optimiza-tion-based separation with deep neural refinement to achieve accurate anomaly detection with listenable explanations. Anomaly detection requires no predefined anomalous categories, and understanding the basis for each detection helps security personnel assess the validity of each detection. End-to-end anomaly detectors can achieve high detection accuracy but output only scalar scores, and existing anomalous sound separation methods provide listenable explanations with limited detection accuracy. AnomSep first obtains preliminary normal-sound and anomalous-sound estimates through optimization-based separation using a dictionary learned from normal sounds in the target environment. Deep neural refinement then improves the preliminary estimates in a pretrained audio latent space. The separated anomalous sound serves as the listenable explanation, and its energy determines the anomaly score. Experiments on synthetic mixtures and real mixtures show improved separation quality over baseline separators. AnomSep achieves detection performance competitive with the end-to-end baseline and provides listenable explanations directly linked to each detection decision.

Federated learning (FL) enables collaborative model training across distributed clients without sharing raw data, making it attractive for privacy-sensitive and bandwidth-constrained applications. A practical challenge in deploying FL is reducing communication overhead while maintaining model performance under heterogeneous, non-IID data distributions. Selecting which subset of clients participates in each training round offers a principled way to address both challenges simultaneously, yet its impact on data diversity and representativeness remains underexplored. In this paper, we investigate three client selection strategies, Power-of-Choice, Fisher information, and Centered Kernel Alignment, evaluating them not only by accuracy and convergence, but through the lens of participation diversity using the Gini coefficient and KL divergence. Under non-IID conditions, restricting participation to 15% of clients per round achieves significant communication savings, and when selection promotes diverse and balanced contributions, it also improves generalization over biased alternatives. This work provides a systematic and reproducible analysis demonstrating that diversity-focused client selection renders communication efficiency and model quality complementary rather than competing objectives.

Recent audio embedding models learn representations that transfer across clip- and frame-level tasks, yet they operate almost exclusively on single-channel signals and are blind to spatial structure. Spatial audio models capture that structure but rely on clip-level supervision or contrastive objectives, yielding embeddings unsuitable for frame-level tasks like sound event detection and localization (SELD). We argue this dichotomy is not fundamental: patch-level spatial self-supervision can jointly learn acoustic and spatial features at both temporal granularities. We introduce AudioSphere, a multi-channel masked autoencoder trained on Ambisonics recordings that reconstructs both spectral content and directional cues via a joint spatiotemporal masking strategy that masks both time-frequency patches and intensity vectors. AudioSphere supports frame-level predictions while producing strong clip-level representations via aggregation. Alongside the model, we release RealSELD, a standardized benchmark for evaluating frozen backbones on real-world multi-channel SELD. AudioSphere enables, for the first time, SELD with a frozen self-supervised backbone, while matching or exceeding prior self-supervised audio embedding models on clip-level classification tasks spanning environmental sound, speech, and music, and on frame-level sound event detection. All code and data are available at https://github.com/labhamlet/AudioSphere and https://huggingface.co/labhamlet.

Large Language Models (LLMs) show remarkable performance based on parametric knowledge, i.e., their ability to recall facts and associations internalized in a model's weights. However, studying this capability on established multi-hop benchmarks may be complicated by benchmark contamination, as many of these benchmarks have been publicly available long enough to plausibly appear in modern pretraining corpora. As a result, correct answers under missing evidence may reflect either general parametric knowledge or memorization of benchmark-specific instances. To study how models use internal knowledge when retrieval evidence is incomplete, we introduce FondueQA, in which every required multi-hop fact is uniquely localized to a designated evidence document. This enables controlled document-level evidence ablations to study parametric compensation under incomplete retrieval. Our empirical evaluation reveals that parametric compensation strongly depends on both entity popularity and the structural role of missing evidence. Generally, models compensate more successfully for facts involving popular entities. On FondueQA, we show that bridging evidence can be more useful than the answer-bearing document itself. Moreover, models may recall a missing fact in isolation but fail to use it under partial evidence, particularly when the context contains plausible competing entities, which can reduce Exact Match by up to 15 percentage points.

Phishing website detection remains a critical cybersecurity challenge. Although machine learning classifiers achieve high accuracy on benchmark datasets, they suffer from a fundamental limitation: once trained, their decision boundaries are fixed and cannot incorporate new contextual evidence without retraining the model. This work presents \textbf{PHISHREV+}, a hybrid framework that instantiates the notion of \textit{post-hoc defeasible machine learning}. This framework does not replace the classifier. It treats the initial classifier predictions as defeasible beliefs and revises them through an Answer Set Programming based non-monotonic reasoning(NMR) layer. The framework operates in two phases. In the machine learning phase, four classifiers (SVM, KNN, DT and RF) are trained on the 87 feature benchmark dataset. In the post-hoc NMR phase, four defeasible rules derived from exploratory data analysis of the training set, and validated through Chi-square and Mann-Whitney U tests, formally revise initial predictions. The proposed framework achieves consistent false positive reduction across all our classifiers SVM (41$\rightarrow$39), KNN (47$\rightarrow$46), DT (62$\rightarrow$61), and RF (52$\rightarrow$51) while maintaining complete false negative parity with zero increase overhead. A comprehensive ablation studies across seven rule configurations confirms that the hierarchical combination of all four defeasible rules is necessary to achieve this optimal balance. Instance level explainability is provided through eXplainable ASP (XASP), which generates formal derivation graphs showing which defeasible rules fired and why the initial belief was retracted, a capability absent from all existing phishing detection approaches.

Weight-space programming sets every weight of a feed-forward network by hand, from an explicit specification of the desired computation, with no gradient descent and no training data anywhere in the construction. We ask how far a large language model can carry this practice, zero-shot, across games of deliberately increasing difficulty. We design a four-rung ladder (Nim, tic-tac-toe, two sizes of Connect Four, and a chess endgame) spanning four axes of difficulty, and build a second, independent network at each rung, using a structurally unrelated game at matched difficulty, to separate what a rung's finding says about the difficulty level from what it says about the one game used to measure it. Ten networks are built and independently verified. Six admit a complete, non-statistical proof of correctness against an independent oracle. The chess-endgame rung is exact where correctness is cheap, move legality and, in one network, an exact one-ply mate detector, and collapses where it is expensive, composing a multi-move mating plan, from 84.1-100% success at the one-ply tactics it was built to get right down to 0.57-9.13% in full adversarial play; its replicate shows this collapse is an order of magnitude less severe under a shorter required plan and exact rather than heuristic safety checks. The standard-sized rung confirms, by holding hand-set evaluation weights fixed and adding only shallow search, that the remaining gap is an architectural ceiling rather than a matter of engineering effort, and this exact collapse reproduces across two structurally unrelated games. We release all ten networks, their independent oracles, construction-cost measurements, and a cold-prompt protocol for a future cross-model comparison

The ENTSO-E Transparency Platform publishes generation, load, and cross-border power flows for European bidding zones. The values carry undocumented errors and violate energy conservation, and data reconciliation consists in correcting them to satisfy that balance. In this work, we formulate data reconciliation as Bayesian inference. Classical methods implicitly take a uniform prior, which we replace with a diffusion model estimated from the observations alone by empirical-Bayes expectation-maximization. The resulting posterior is sampled with a guided process whose projection step makes every sample satisfy the balance exactly. On synthetic systems of six to thirty-five bidding zones, built from ENTSO-E resource-adequacy data, the diffusion prior gives the most accurate reconciliation, increasingly so as the system grows. Its credible intervals are calibrated or conservative, and it remains the most accurate under a misspecified noise covariance. Finally, when the generation of entire zones goes unreported, the classical methods fail while the learned prior recovers it from cross-zone structure.

LLM-based agents increasingly automate business and consumer tasks but are slow and expensive. Decision models, a new class that chooses among options without generating text, are far faster and less costly. Developers call them "System One models," referencing dual process theory. We ask whether such models can be steered by digital nudges online and whether, as the theory predicts, automatic nudges affect them more than reflective ones. We adapted an online shopping experiment, previously run with humans and LLM agents, to three decision models (Jev, Decision-1, d1). Across 600 runs per model, they chose a preselected default in 97.5-99.9% and a product with social influence information in 75-96% of choices, versus 47-49% in the control. All responded more strongly to the automatic nudge, the opposite of what has been reported for reasoning-capable LLM agents. We provide first evidence that non-generative decision models are susceptible to digital nudges in a System 1 pattern, implying agent-type-specific nudge effects.

Direction of arrival (DoA) estimation, a classical signal processing problem, has experienced renewed interest with the advent of integrated sensing and communication (ISAC). While decades of research have yielded highly accurate DoA estimators, hardware impairments have frequently been neglected, despite degrading both DoA estimation accuracy and downstream communication performance in ISAC systems. This work introduces a joint array calibration and DoA estimation method based on the simultaneous orthogonal matching pursuit (SOMP) algorithm, following the model-based machine learning paradigm. The proposed method enables the calibration of antenna arrays subject to severe gain-phase and antenna location impairments through projected gradient descent (PGD), with either supervised or unsupervised objectives. Additionally, DoA estimation is enhanced by extending SOMP to continuous dictionaries via PGD refinement steps. Simulation results demonstrate that the proposed method enables efficient hardware impairment compensation, even without labeled samples, while outperforming conventional sparse recovery and subspace-based DoA estimation baselines.

In the age of digital transformation, companies are increasingly relying on specialized software to manage the complexity of projects. However, when selecting tools, features and costs often take precedence over user experience (UX), leading to low adoption rates, user frustration, and project delays. This study empirically evaluates the user experience of leading project management tools (Jira, Trello, Monday.com, Asana, and Microsoft Project) across various project scales using a mixed methods approach. The System Usability Scale (SUS) was used as the primary quantitative instrument, with data collected from 141 experienced professionals via online questionnaires. The results reveal a significant trade-off between functionality and usability. Trello achieved the highest SUS score (77.5), indicating a superior user experience for small projects and agile teams, while Microsoft Project scored the lowest (51.2), suggesting significant usability challenges despite its robust features. Jira demonstrated a balanced profile, offering powerful features with manageable complexity. Statistical analyses confirmed that the choice of the optimal tool is highly context dependent, with team size and project complexity influencing UX outcomes. The findings underscore that UX is not a secondary feature but a critical success factor for software acceptance and project effectiveness. This study provides an evidence-based framework for aligning tool selection with the project context, offers practitioners actionable insights, and identifies design opportunities for software developers.

Background: Binary hypotension forecasts do not describe the depth and duration of future low pressure. Burden forecasts may support graded early warning, but better forecast scores need not yield more useful alerts. Objective: To evaluate conditional quantiles of the next 15-minute mean arterial pressure (MAP) deficit below 65 mmHg and distinguish forecasting performance from episode-warning performance. Methods: Public VitalDB numeric arterial MAP was reduced to 10-second medians without forward filling. Subject-disjoint training, validation and test splits included 70,402, 14,270 and 15,575 labeled windows. We compared climatology, calibrated persistence, last-MAP-only boosting, and direct and residual quantile boosting, with and without additional MAP-history features. Patient-weighted pinball loss and subject-bootstrap intervals evaluated forecasting. Validation selected alert thresholds before a hash-recorded test evaluation. Results: In the test split, the prespecified residual-versus-persistence comparison at the 90th percentile yielded losses of 2.256 and 2.900 mmHg-min, respectively (difference -0.644; 95% interval -0.814 to -0.487). Direct prediction achieved 2.196, and direct prediction with additional MAP features achieved 2.193. Episode-warning results were sensitive to the alert rule and operating threshold. Conclusions: This retrospective benchmark separates burden-forecast skill from warning-policy performance. Findings support further study of direct burden estimation; they do not establish patient benefit, prospective reliability or transportability beyond the source hospital.

Cooperative perception enables an ego agent to acquire information from collaborators beyond its own sensors' reach. Yet, the most beneficial messages are often the hardest to verify from local observations. Pretend Benign (PB) exploits this asymmetry by targeting regions with weak local verifiability and applying multi-objective perturbations, making harmful feature changes appear to be benign assistance. However, PB is victim-dependent: selecting regions and optimizing perturbations require the victim model's architecture, parameters, and gradients, limiting transfer to unseen victims. A transferable attack must capture properties shared across cooperative detection models rather than those specific to one network. We address this by building a victim-agnostic attack from a heterogeneous ensemble of surrogate models and encoding its capabilities in a reusable dictionary to reduce per-message generation cost. Specifically, Transferable Pretend Benign (TrPB) transfers PB's region selection criterion and multi-objective perturbation strategy across heterogeneous surrogates to construct an offline attack. We further propose RT-TrPB, which encodes transferable regions and perturbation patterns into a region predictor and reusable perturbation dictionary, amortizing per-message attack generation. Experiments on widely used cooperative perception benchmarks show that RT-TrPB bypasses the consensus-based defenses ROBOSAC and FLD with success rates approaching 100% and generates each perturbation in under 56 ms. To counter PB-style attacks masquerading as benign collaboration, we also propose STTrust, which verifies received messages using spatial and temporal consistency and incorporates identity-level trust. STTrust detects over 90% of PB-style attacks and restores perception performance to near the attack-free level.

Mixture-of-Experts (MoE) training is often not stationary: the data mixture changes at curriculum boundaries, when a model is continually pretrained into a new domain, or when the corpus is refreshed. Routing is learned against the current mixture, so each such change disrupts the per-expert load the router had converged to, and existing balancing methods have only been evaluated on stationary data. We rethink load balancing as a feedback control problem, in which the deviation of each expert's load from its fair share is the control error and the per-expert routing bias is the control input, and define a family of controllers, the PID-Balancer, from the proportional, integral, and derivative terms of classical PID control. Two widely used methods belong to this family: the auxiliary balancing loss behaves like a proportional controller, and the loss-free bias update is an integral controller acting on the sign of the error. Reading the router as a static map with a one-step delay, we show that the integral term is the only one that both removes persistent imbalance and averages out the step-to-step fluctuation of the measured load, whereas the proportional and derivative terms pass or amplify that fluctuation into the routing bias. We therefore adopt PID (I), the pure-integral member of the family with a single gain. Two settings are evaluated: (i) training an OLMoE backbone from scratch with 16 to 64 experts per layer, through a mid-run mixture switch, and (ii) continual pretraining of the released OLMoE-1B-7B on a new corpus. PID (I) holds balance before the shift and restores it more tightly afterwards than either the auxiliary loss or loss-free balancing, at comparable perplexity.

SRAM's power-performance-area-stability trade-off depends on bitcell topology, a discrete architectural choice that interacts non-separably with continuous transistor sizing. However, no open-source SRAM tool treats topology as an optimizable variable. OpenRAM generates SRAM layouts but performs no optimization, while OpenYield supports high-fidelity evaluation but restricts optimization to scalar sizing of a fixed 6T cell. OpenYieldV2 closes this gap through three integrated contributions. First, a fully parameterized compiler generates complete 6T and 10T SRAM macros with peripheral circuits and replica-bitline self-timing, enabling topology comparison under identical simulation conditions. Second, a few-shot TabPFN surrogate learns topology-dependent PPA responses with limited full-macro SPICE samples. Third, a coarse-to-refined multi-objective optimizer combines global exploration of discrete architectural choices with local refinement of continuous transistor sizes. Experiments across varying design conditions reveal a topology decision boundary that shifts with operating constraints rather than following a fixed design rule. Jointly optimizing across this boundary improves hypervolume by up to 17.3\% over the strongest single-topology baseline. OpenYieldV2 is open-source to support reproducible SRAM design research.

Gaussian processes on graphs provide a principled framework for prediction and uncertainty estimation, but modelling heterogeneous networks often requires balancing flexibility against parameter complexity. Variability at a node may reflect useful signal that should propagate through the network, unreliable measurements whose influence should be suppressed, or a mixture of both. We introduce Variance-Coupled Gaussian Processes (VCGPs), which use a positive auxiliary signal to modulate graph interactions through a single learned coupling exponent. The exponent controls how strongly high-variability nodes propagate or attenuate information, while a cluster-wise extension accommodates different regimes within the same graph. Under stated structural assumptions, we characterize the unique stationary-rescaling point of the basic family, where correlations match those of a stationary graph kernel while node-dependent variances remain. Experiments on traffic, air-quality and weather datasets show competitive prediction and uncertainty estimation relative to stationary and per-node non-stationary baselines, with the largest gains on networks requiring multiple coupling regimes. Stationary correlations after variance rescaling remain sufficient on some data, and performance depends on the auxiliary signal and model selection.

Multimodal dynamic graphs are pervasive in real-world applications, where nodes and edges are accompanied by multimodal signals, including text and images. Despite their practical importance, benchmarks and methods that explicitly study multimodal learning in dynamic graphs remain scarce, hindering the full understanding and effective utilization of multimodal information for downstream tasks. To address this gap, we introduce MM-DyGraph, to the best of our knowledge, the first benchmark for multimodal dynamic graph learning, consisting of diverse real-world datasets with fine-grained temporal graph evolution and rich multimodal attributes. MM-DyGraph provides three evaluation tasks to comprehensively assess model performance in dynamic settings. We benchmark representative state-of-the-art dynamic graph methods and reveal that naive feature concatenation yields target-level inconsistencies: multimodal inputs benefit certain prediction targets while harming others. This observation indicates that the modality informativeness depends on the specific prediction target over time. Building on these observations, we propose a Target-aware Spatial-Temporal Multimodal Fusion (TST-MF) framework, which jointly conditions multimodal fusion on the prediction target and spatial-temporal relations, enabling context-adaptive cross-modal interplay for target-specific representations. Extensive experiments over diverse datasets and tasks show that TST-MF delivers robust and competitive performance, achieving clear gains in multimodal settings.

Diffusion language models (DLMs) enable parallel token generation through iterative denoising, yet their inference remains costly due to repeated attention over a growing historical prefix and repeated processing of the entire current generation block. Historical attention becomes increasingly expensive as generation progresses, while all current-block positions are propagated through the Transformer at every denoising step even though only a small fraction are typically resolved at that step. In this paper, we introduce Duet, a training-free dual-sparse inference framework that addresses these challenges through prefix Key-Value (KV) sparsity and early query sparsity, respectively. Specifically, Duet exploits the concentration and cross-step stability of historical attention within each generation block to construct a compact historical KV subset that is reused across denoising steps. For query sparsity, Duet predicts update positions from shallow representations and propagates only the corresponding queries through the suffix Transformer layers, avoiding deep computation at positions unlikely to be updated in the current step. Experiments show that Duet maintains competitive accuracy while achieving 1.34$\times$/2.83$\times$ end-to-end/per-step speedups on LLaDA2.1-mini and 2.41$\times$/4.72$\times$ on SDAR-8B-Chat-b32 at a 32K context length.

Controllable molecule generation is crucial for diverse scientific applications, such as drug discovery and materials design. While large language models (LLMs) show great promise, their dense and entangled representations impede precise control over the generation of molecules with bespoke substructures or properties. To address this, we propose Sparse Representation Editing (SpaRE), an interpretability-driven framework for fine-grained and precise control in LLM-based molecule generation. The crux of SpaRE is to learn an overcomplete sparse feature space that disentangles LLM representations into a compact set of latent features corresponding to chemically meaningful concepts. Within this space, we can directly manipulate these concept-aligned latent features to achieve (1) local control, by generating target atoms and functional groups at specified positions; and (2) global control, by customizing the overall structural and physicochemical properties within defined ranges. In this way, our framework advances interpretability from post-hoc analysis to actionable generative control. Experiments show that SpaRE can generate chemically desirable molecules under complex constraints in real-world scenarios, while offering mechanistic insights for quantitative structure-property analysis. The code and demo are available at https://github.com/WanyuGroup/ICML2026_SpaRE.

Model-Based Systems Engineering (MBSE) is a promising approach to improve systems engineering practices, particularly by enhancing requirement management, traceability, and early validation and verification (V\&V). However, its adoption remains challenging, especially for small and medium-sized enterprises (SMEs), which often lack structured processes and solutions adapted to their scale and constraints. Our goal is to propose a lightweight MBSE methodology tailored to the needs and constraints of SMEs. This paper focuses specifically on requirements refinement, traceability, and early V\&V. The approach integrates MBSE principles with artificial intelligence-assisted analysis to improve requirements quality while remaining compatible with existing tools and practices. A proof of concept is developed and applied in an industrial case study to evaluate the feasibility of the approach. The results demonstrate how early V\&V activities can be supported within an SME context, while improving traceability across engineering artifacts. These findings highlight the potential of a lightweight, integrated MBSE approach to support the progressive adoption of MBSE in SMEs, thereby improving overall engineering efficiency.

Cross-Layer Transcoders (CLTs) decompose the activations of large language models into human-interpretable features and trace how these features interact across layers to form circuits, making them a key tool for mechanistic interpretability. However, training CLTs effectively and efficiently is hard. Beyond the usual trade-off between sparsity and reconstruction quality, a CLT has to model all layers at once, which adds a further challenge. The common JumpReLU recipe controls sparsity only indirectly through a penalty, while the direct extension of TopK selects a fixed number of latents in every layer. Yet reconstruction difficulty differs widely across layers, and intermediate layers are much harder than others, so giving every layer the same budget is suboptimal. We propose Cross-Layer TopK, which sets a single L0 budget for the whole model and dynamically allocates it across layers during training. In a comparison with JumpReLU and other TopK variants on Qwen3-8B-Base, this gives lower FVU while using 20% fewer active latents than JumpReLU. The better reconstruction carries over to circuit tracing, yielding more complete circuits with higher replacement and completeness scores. The fixed per-token budget of Cross-Layer TopK also makes it easy to pair with sparse decoding, which cuts both computation and communication and gives a 2.9x speedup per training step. We further show that Cross-Layer TopK follows scaling laws in both latent size and sparsity. We have open-sourced the weights of our CLT at https://huggingface.co/ByteDance-Seed/Cross-Layer-TopK-CLT.

Artificial Intelligence (AI) is increasingly embedded in high-risks decision-making systems across domains such as healthcare, finance, and the judicial system. As reliance on predictive models grows, so does the demand for Explainable AI (XAI) to ensure transparency, trust, and interpretability. However, existing XAI approaches often fail to present explanations in a format accessible to domain experts, limiting their practical utility. Large Language Models (LLMs) have recently emerged as a bridge between complex models and end-users, offering natural language explanations. Yet, their use introduces new risks, including hallucinated outputs and lack of source traceability. To address these challenges, we propose a Clinical Decision Support System (CDSS) that not only combines standard predictive models with XAI techniques but also incorporates a Retrieval-Augmented Generation module as an interactive enhancer. This module grounds explanations in verified medical knowledge, reformulates technical outputs into a human-like format, and enables users-such as clinicians-to actively query the system. By fostering an interactive, human-in-the-loop environment, our approach empowers domain experts to explore model decisions, contextualize explanations, and build trust in AI-assisted diagnostics. The study concludes with preliminary experiments validating the proposed methodology.

Low-precision matrix throughput on modern GPUs is growing far faster than vector-arithmetic throughput. Attention therefore cannot realize the available GEMM performance because softmax has become the pipeline bottleneck. We introduce Fused-Softmax Attention, a training-free algorithm that reformulates and approximates softmax to fold its dominant operations into the attention GEMMs, moving work from vector units to matrix cores. Implemented for the AMD CDNA 4 architecture, our MXFP8, MXFP6 and MXFP4 kernels reach approximately 3.0, 4.0 and 4.6 PFLOP/s on an AMD Instinct MI355X GPU, yielding speedups of $2.1\times$, $2.8\times$ and $3.2\times$ over the BF16 attention. Importantly, the additional error introduced by Fused-Softmax remains below the error floor set by quantizing the GEMM operands. On Wan2.2-T2V-A14B, a hybrid MXFP6/4 schedule delivers up to $2.05\times$ end-to-end speedup while preserving the visual quality of the outputs. The kernels ship in AMD's open-source AITER library, and they powered AMD's MLPerf Inference v6.1 Wan2.2 submission, which ran $1.7\times$ faster than AMD's previous-round result on the same hardware. Our results show that redesigning softmax to utilize the matrix pipeline enables attention to capitalize on continued growth in low-precision matrix throughput.

Automated fish feeding intensity assessment (FFIA) is important for precision aquaculture and feed cost optimization. However, unimodal methods are vulnerable to acoustic noise and visual disturbances, while existing multimodal frameworks often rely on computationally intensive 3D convolutions or Transformers, limiting their use on edge devices. To address these limitations, we propose HMMF-Net, a lightweight multimodal architecture that combines early multi-frame visual encoding and multi-axis acoustic modeling through a hierarchical bidirectional messenger cross-modal fusion (BMCF) mechanism. The visual branch captures temporal variations in feeding behavior by stacking temporally distant frames as a multi-channel input to an adapted 2D EfficientNetB0 backbone. The acoustic branch employs adaptive multi-axis residual units (AMARU) to extract temporal, frequency, and local time-frequency representations from log-mel spectrograms. BMCF progressively exchanges information between the two modalities at intermediate representation levels through compact messenger tokens, reducing the computational overhead of dense cross-modal attention. Evaluated on the AV-FFIA dataset using five-fold cross-validation, HMMF-Net achieves 96.77% accuracy and 96.88% F1-score on the held-out test folds, with 3.987M parameters and 1.002 GFLOPs. Compared with the best-performing visual configuration (95.49% accuracy), HMMF-Net reduces the classification error by 28.38%. Deployment on an NVIDIA Jetson Orin Nano using TensorRT with FP32 precision maintains comparable accuracy, with an inference time of 11.54 ms per 2-second sample and a model size of 18.81 MB. These results demonstrate that HMMF-Net combines effective multimodal representation learning with computational efficiency, enabling accurate and practical FFIA on resource-constrained edge platforms.

Existing Korean benchmarks for Large Language Models (LLMs) evaluate performance across diverse knowledge domains, but their accuracy scores cannot distinguish whether incorrect responses arise from a lack of required knowledge or a failure to apply that knowledge correctly. Our analysis of existing Korean benchmarks reveals that most questions can be answered through factual recall alone, while those demanding greater application also differ in required knowledge, making it difficult to isolate application ability. To address this, we introduce KLEAR, a framework that constructs knowledge-matched contrasts by varying application demand while preserving the required knowledge. During construction, KLEAR identifies the knowledge underlying each source question as an explicit anchor bundle to guide generation of questions requiring more complex application. Generated candidates then undergo automated and human validation, with iterative revision based on evaluation feedback. Experiments demonstrate that the resulting knowledge-matched sets largely satisfy the intended constraints, enabling controlled evaluation across diverse LLMs. These evaluations reveal that increasing application demand exposes failures that persist even when the required knowledge is provided.

Benchmarks for open-ended generation often rely on large language model (LLM) judges, whose reliability in low-resource languages is uncertain. We introduce Khmer PFGen-Bench, to our knowledge the first judge-free benchmark for multi-sentence answers to questions composed directly in Khmer, a low-resource language. Khmer PFGen-Bench adapts the scorer of PFGen-Bench, a benchmark for Japanese that scores responses with keyword rules and n-gram statistics from LLM-generated references. This scoring requires references that are both valid and varied, which LLMs may not produce in Khmer, so we examine how to build such references. We find that post-trained models can give more valid but less varied references than their base variants, and that mixing model families increases variety for the same number of references. Using references from post-trained models in nine families, the benchmark yields model rankings that are stable across reference sets and consistent with a native speaker's ratings. These findings offer guidance for building judge-free benchmarks in other low-resource languages.

Automated harness optimization improves agents by learning from trajectories and task feedback. Existing offline methods use reference answers to optimize harnesses for reuse on held-out tasks. However, many real-world tasks are open-ended, while long runs expose new difficulties that offline-optimized harnesses may not address. To study harness optimization in these settings, we introduce PerfRace, a benchmark for long-running performance optimization of software projects, and investigate whether agents can use task-specific experience to recursively improve their harnesses during an ongoing task. To this end, we introduce BOXBOX, which enables agent-triggered harness revision and kernel-managed hot reloading while preserving execution context. We evaluate BOXBOX on PerfRace under matched test-time budgets, comparing it with Frozen Harness, offline Harness Evolution, and Harness Scaling. BOXBOX achieves higher speedup with more efficient use of task time and model cost, reaching a Speedup@4h of 1.70x, compared with 1.54x for the strongest baseline.

Estimating the state of a dynamical system from observations is a crucial task in robotics and control. Despite their empirical success, recurrent models often lack formal guarantees on convergence and robustness. We bridge this gap by analyzing these architectures from the perspective of nonlinear observer theory. By framing RNNs as a time-discretization of continuous latent dynamical systems (Neural ODEs), we demonstrate that enforcing a contraction condition on the latent flow guarantees the existence of a smooth, injective map of the true physical state into the latent space. We prove that this architecture acts as a tunable observer: by increasing a single scalar gain that scales the latent vector field, the estimation error can be driven to an arbitrarily small neighborhood of zero at an exponentially fast rate. We further establish quantitative bounds on the MLP complexity required to decode this latent state, and extend our formal guarantees to the discrete-time residual RNN implementation. We validate the theory on a real mobile robotic task, where a provably contracting model outperforms a vanilla GRU with substantially more stable training. We also report that unconstrained GRUs converge to contracting dynamics on this task, suggesting that contraction is an emergent property of well-trained sequence models.

Iterative repair returns a failed program's error to a large language model (LLM) and asks it to fix it. After iterative repair, the remaining failures are mostly assertion failures: with four Qwen2.5 models, they make up 79% of residual failures on BigCodeBench, all on HumanEval and 93% on LiveCodeBench. We argue, through an information view of the feedback, that an assertion failure says little about where the fault lies. We propose Breakpoint Debugging (BPD), which returns the failed program annotated inline with its variables' runtime values. We take the tasks that iterative repair failed to fix with five repair attempts and compare two ways of spending five further generations on them: BPD, and resampling. On BigCodeBench, BPD repairs 26.7% of these tasks against 22.4%, and its advantage grows with model size. On HumanEval, BPD outperforms resampling but only slightly, and on LiveCodeBench, fresh sampling repairs more (6.5% against 4.9%). Furthermore, BPD reduces unchanged-code resubmissions from 28% to 8% on BigCodeBench. Using only the first failing test's trace, it outperforms tracing all failing tests while using fewer tokens than both all-trace BPD and plain iterative repair. On LiveCodeBench, BPD gives no gain even on the public tests, so a visible failing input is not enough for a trace to help. These results match the information view: a trace helps only when the test exposes a failing input, and the fault is an implementation slip; otherwise, a repair loop should resample rather than trace.

Reinforcement learning (RL) is now central to training Reasoning Language Models (RLMs), but RLM pipelines are difficult to scale and costly to run because each RL step couples multiple large models, heterogeneous stages, cross-step state, and complex asynchronous dependencies. Existing systems typically optimize only parts of this design space. We address this complexity by jointly optimizing the pipeline structure as well as its execution using a novel data-centric graph intermediate representation called RLM Pipeline Execution Graph (RPEG). RPEG exposes structural choices such as bounded asynchrony and actor-critic sharing together with execution optimizations such as stage fusion, fine-grained scheduling, and multi-dimensional model parallelism. It is coupled with a comprehensive analytical performance model for RLM training, jointly capturing compute, communication, memory, and hierarchical rack-, node-, and GPU-level effects. Across systems with up to 256 GH200 GPUs, our distributed runtime engine implementing RPEG improves performance by up to 2x over ReaL and reduces step latency by up to 2.48x over verl, while increasing hardware utilization by up to 30%. Overall, RPEG provides a graph-driven foundation for next-generation training pipelines for frontier reasoning LLMs.

In neuroscience, representational drift is a change in which neurons carry a representation while behaviour stays the same; we measure its analogue in OpenCoder-1.5B during pretraining, for 169 Python constructs over two trillion tokens at a constant learning rate, with a noise unit on every change, at checkpoints as close as 1,000 training steps. For 16 of the 169, those whose grammar fixes a next token, the loss at those positions reaches its floor within the first 4% of training, while the model's reliance on the neurons carrying them grows to the end: removing them costs 0.06, 0.25 and 0.40 nats more than removing a random set, at 4 billion, 84 billion and 2 trillion tokens. Which neurons carry a construct never settles: for all 169 the carrier set is still changing at the last checkpoint, and an abrupt settling at any date from 84 billion to 1.5 trillion tokens, or a gradual one before 500 billion, would have been detected in 99% of injected cases. The set measured at 84 billion tokens, removed at the end, cannot be separated from a random set, so the function is carried by whichever neurons are current. Up close the change is a persisting pool and a changing sample: a neuron that leaves a construct's set returns at three times the rate of a matched neuron, and no set reassembles. The hypothesis we fixed in advance, that the carriers settle, is refuted; we discuss candidate causes and name the experiment that would separate them.

Many news reports of a claim are copies of one source: wire stories, content shared within a media group, syndicated features and press releases. We ask whether LLM judges and continual learners treat such copies as independent confirmations. In 200,000 English CC-News articles, a MinHash detector links 18.9% to a near-copy from another outlet (precision 0.97 once boilerplate is removed); it links 92% of edited reprints to their originals but none of the LLM rewrites by phi4, Llama 3.2 or Gemini 2.5 Flash. Shown three identical reprints of a claim about an unknown person instead of one report, Llama 3.2 3B and phi4 14B raised the share of lies they accepted from 0.06 to 0.71 and from 0.21 to 0.90 and ignored a note that the reports came from one source; Gemini 2.5 Flash already believed a single report, and Gemini 2.5 Pro followed provenance notes and tables. In a simulated continual-learning stream, showing the judge one report per detected chain cut the copied lies it accepted by 0.10-0.24 across four judges, at a mean cost of no more than 4.2 percentage points of accepted truths. Against reworded copies, a test of how many of a report's named entities are absent from an earlier look-alike separates LLM rewrites from 100 labelled independent reports of the same event with AUC 0.98-0.99 and wrongly flags one; entity overlap alone does as well, TF-IDF and a current embedding model come close, and the test catches few natural same-source derivatives. With the miss patterns of generated rewrite networks, it cut the reworded-copy lies accepted from 0.67 to 0.20 (0.40 when the rewriter invents details). Counting reports or chains alone is not a reliable defence, the effect of provenance information depends on the judge, and colluding sources remain the open problem. Code and labels: https://github.com/Metalesaek/copies-experiment

We study Vietnamese statutory Article retrieval on DRiLL using features for provision identity and local form, same-law support, and citation and lineage evidence. Across five random seeds, selectors sharing candidates, listwise Qwen evidence, and learner settings improve official $F_2$ from 71.77% to 72.91%, with the largest descriptive gain on two-Article questions. Added features account for 18.20% of the proposed selector's total decision-tree split gain, measured as training-objective reduction, mainly from same-law and citation evidence. Cases reveal base provisions and amendments, companion recovery, distractor removal, and failures. These findings motivate research on other datasets, representations, and learners without establishing individual feature effects or requiring LightGBM, hand-crafted scores, or explicit graphs.

Shielding is a safe reinforcement learning technique that provides safety guarantees while learning, by blocking unsafe actions. Traditional shielding relies on a complete model of the environment to provide long-term guarantees. Probabilistic logic shielding (PLS) lifts the requirement of having such model, but only considers safety in the short term, typically looking only one or few steps ahead. We propose Probabilistic Option Shielding (POS), which extends PLS to the option level, recovering long-horizon safety without reintroducing the need for a complete environment model. POS learns a model-free critic directly from one-step PLS signals, to estimate the probability that an entire option will complete safely under uncertainty. We evaluate POS on safe navigation tasks in the LogiCity environment, and show that it achieves considerably safer training while matching the task performance of standard step-level PLS and other safety related reinforcement learning baselines.

Interpretability has largely moved from reading a model's neurons directly to reading a second model trained on the first: a probe, a sparse autoencoder, a crosscoder. The move answered real faults -- mixed meanings, thresholds without a unit, readings that did not transfer from one study to the next -- but gave up what a direct reading alone provides: nothing trained in between, the neurons themselves named, and a reading that can be repeated and compared. Here the direct reading is made a measurement. A neuron's activation varies with the prompt, so its reading is a distribution over a population of stimuli. We declare that population: every factor around a target construct is drawn at its rate in a reference corpus and logged. From the declared distribution we generate a battery of prompts, and draw it twice. From the battery we compute the calibrated concept profile, one number per neuron, stated in units of the neuron's own spread on ordinary input; the two draws give every number a noise unit. Before the profile is read, a fixed set of checks licenses the battery on the model, and the smallest change the profile can resolve across training checkpoints is measured. Realised for Python on OpenCoder-1.5B and followed across checkpoints spanning two trillion training tokens, the profile gives what the alternatives do not report: a unit on every number, a chance level for every comparison, and which neurons carry a construct at every checkpoint. The battery, the code and the data behind every result are released.

Patent analytics increasingly relies on transformer representations and large language models (LLMs), yet linking conceptual classification, topic formation and interpretation requires explicit decisions about model roles and the evidence retained at each stage. In this design-oriented study, we propose and demonstrate a patent analytics pipeline that links these tasks and keeps each decision open to audit. Three LLMs separately classify each patent by novelty (radical or incremental) and innovation type (product, process, service). Radical and innovation-type labels each require two supporting votes, with Incremental used as the fallback novelty label. Topics are discovered by clustering PatentSBERTa embeddings, with parameter ranges scaled to corpus size. A patent-specific stopword list is applied only when topics are represented with class-based term frequency--inverse document frequency (c-TF-IDF), and a second embedding model supports inter-topic similarity, hierarchical refinement and visualization. An LLM then labels each topic from its ranked keywords and representative patents. We demonstrate the pipeline on 30,964 public-sector-related patents retrieved from Lens.org. Across five category corpora ranging from 591 to 29,215 patents, the pipeline produced 74 category-specific topic instances with a macro-average document coverage of 98.1% and topic diversity of 0.900--1.000, and fewer than 1% of classification outputs returned UNKNOWN. Because every model vote and the evidence behind each topic label are retained, analysts can audit and revise the outputs. The demonstration establishes the feasibility and traceability of the pipeline; testing its labels against expert coding and benchmarking it against alternative pipelines are identified as the next steps. The pipeline offers a reusable basis for LLM-assisted patent analysis in technology intelligence.

Structural role embeddings describe a node by its position in a graph - hub, bridge, periphery, member of a dense core - rather than by who its neighbours are. Existing methods either learn coordinates specific to one graph (struc2vec, Role2Vec, RESD) or read scales and normalisations from the whole graph (GraphWave, Digraphwave), so the same neighbourhood can receive different coordinates in two graphs; few read edge directions. We present RoleWalk, a training-free embedding that follows a short lazy random walk from each node and records, at each step, two distributions: that of the transition probabilities, through its characteristic function at a fixed grid of frequencies, and that of the walker's distance to its starting node. On directed graphs, walks that go forward then backward add who shares a node's targets or sources. The embedding depends only on a node's neighbourhood within a fixed radius, so its coordinates mean the same thing in every graph. We show that, without curvature, it is the kernel mean embedding of the walk's footprint, that it separates graphs that the Weisfeiler-Leman test cannot, and that pruning small probabilities comes with a per-node error certificate. Empirically, RoleWalk is competitive with the best role embeddings on five classification benchmarks at a fraction of their cost, has the highest scores on real directed roles in Wikipedia adminship, exceeds the published accuracy of Digraphwave at every noise level of its own Enron network-alignment benchmark, transfers centrality predictors to graphs four times larger better than existing inductive embeddings, hand-made features and a trained GIN, and embeds a million nodes in 20 s to 3 min on one CPU core. The code is open source (BSD license): https://github.com/kerighan/rolewalk