In briefOptimal regret bounds for online market making with limit order book feedback
SummaryThe paper studies online market making, where a market maker posts bid and ask prices before seeing the market price and a trader's private valuation, which is revealed only when no transaction occurs. It improves a prior expected regret bound of O~(T^{2/3}) to a high-probability bound of O~(√T). The authors first achieve this under full feedback using two coupled grids with Hedge, then extend it to limit order book feedback. They also show that when both prices and valuations are adversarial, sublinear regret is impossible even with full feedback.
Maria Elena Vischi, Francesco Emanuele Stradi, Alberto Marchesi
In briefNAViLoss: A Navigation-Aware Dual-Residual Loss for Physics-Consistent Underwater Velocity Learning
SummaryThis paper proposes NAViLoss, a robust, uncertainty-aware loss function for learning-based AUV velocity estimation. It jointly penalizes velocity residuals in the navigation-state domain and beam-consistency residuals in the DVL measurement domain. Integrated with DeepONet as the NAVi-DeepONet model and evaluated on about 10,000 m of semi-synthetic sea-trial data, it reports a 44% accuracy improvement over conventional and learning-based baselines.
In briefSilhouette Operator: Recovering Low-Rank Measures from 1D Projections
SummaryThis paper develops a recovery framework for two-dimensional low-rank signed measures, which are finite sums of product measures with one-dimensional factors, using linear "silhouette operators" that map a measure to one-dimensional projected marginals. The authors show that 2k suitably chosen projected marginals identify every compactly supported signed measure of rank at most k, that this number is optimal, and that projection directions cannot be chosen arbitrarily. They also give conditions under which pairwise marginals identify higher-dimensional models and introduce silhouette mixture estimation (SME), which matches projected marginals to empirical ones in Wasserstein distance. Combined with one-dimensional density estimators, SME is reported to perform well against several baselines at moderate dimension and sample size.
In briefSelf-organization in closed systems via constrained geometric radiation
SummaryThis paper studies how dynamic order arises spontaneously in closed systems without external driving. Simulations of coupled metric evolution systems reveal a four-stage cycle of stress accumulation, super-exponential radiation, chaotic collapse, and convergence to a fractal limit cycle, termed a wedge-shaped attractor. The authors identify four jointly sufficient conditions and prove several theorems, and report a critical noise threshold and universality across 12 configurations.
In briefFrom Pareto to Preference: Personalized Test-Time Scaling via Amortized Agentic Policy Discovery
SummaryThis paper introduces Personalized Test-Time Scaling, which seeks executable controllers that maximize joint satisfaction of users' multidimensional requirements for accuracy, latency, and inference cost. It proposes PersonTTS, an amortized framework that reuses prior search experience through requirement-matched initialization and source-distilled procedural guidance, while evaluating every candidate on the target profile. Experiments on AIME and HMMT show it outperforms strong TTS baselines on unseen user profiles and held-out problems, and reduces discovery-agent time and cost under the same candidate-evaluation budget.
In briefSystem Switch: When a fast decision model should stop and think
SummaryThe paper studies dual-process agents in which a fast learned policy makes every decision and hands control to a reasoning vision-language model only when a gate opens, while the game keeps running. Experiments use closed-loop Doom and the open "System One" typed-decision models on 900 held-out questions. Accuracy, calibration and sensitivity are shown to be distinct, and deferring the least confident 30% of decisions gains over random deferral. However, no variant reaches the exit in closed-loop play across 33 games.
In briefContext-aware micro-expression recognition in real conversations: a dataset and hierarchical fusion network
SummaryExisting micro-expression recognition methods analyze isolated video clips and ignore the surrounding dialogue's meaning and sound. The authors introduce TriCoME, a trimodal dataset of visual, audio, and text data from real conversations, pairing each micro-expression with its utterance and preceding context, labeled as external stimulus or internal background. They also propose HFFNet, a Hierarchical Focusing Fusion Network that aligns acoustic and semantic streams, compresses long dialogues, and switches between visual-only and multimodal decisions based on visual confidence.
In briefLow-Cost Theorem Proving for Program Verification via Agent Orchestration
SummaryThis paper introduces CoCo-Prover, a method for cost-efficient theorem proving in program verification. It treats proof search as metalevel decision-making under cost, organizing open goals with two-level proof graphs and using an agentic router to choose specialist agents and configurations for each bounded call. On five Lean 4 benchmarks, the authors report a better success-versus-cost trade-off than baselines, with up to 100% solve rate on two benchmarks and up to 30.9% cost reduction versus the strongest baseline.
Shuangjie Yao, Nikolaus Holzer, Mark Paul Santolucito et al.
In briefCERO: Deciding Where and When to Spend Rollout Budget in RL Post-Training
SummaryThis paper studies how to spread a limited rollout budget across the whole training run for group-relative reinforcement learning, rather than fixing a budget per update. It proposes CERO, an online primal-dual scheduler that adaptively chooses which prompts to admit, how often to revisit them, and how many response groups to generate each round. Under matched training-response budgets, CERO achieves the highest avg@16 macro-average on three backbones across five mathematical reasoning benchmarks.
In briefA multi-source ultrasound benchmark showing limits of current self-supervised anomaly detection
SummaryThis paper introduces SADUSI, a multi-source ultrasound dataset spanning many anatomical regions, views, and acquisition protocols, for training and evaluating anomaly detection. The authors evaluate representative self-supervised methods and find they struggle. Reconstruction-based diffusion methods (AnoDDPM, DeCo-Diff) reach pixel-level AUROC of 0.56 to 0.72 and maximum F1 of 0.10 to 0.26, while PatchCore variants perform better at 0.76 to 0.83 AUROC and 0.14 to 0.40 F1.
Marco Riedenauer, Daniel Kienzle, Pratik Mayekar et al.
In briefGauss-Newton curvature accuracy versus indefinite Hessians in low-cost sets
SummaryThe paper studies how accurate Gauss-Newton curvature is for ridge-regularized nonlinear least squares. It proves that, within a low-cost set, every global minimizer has relative Hessian error below (1+√2)/8, while the same set also contains a point with an indefinite Hessian whose relative error is at least 15/8. The authors also give a uniform positive ridge cap and a pointwise certificate controlling the Hessian correction.
In briefInsClaimBench: A Benchmark for Insurance Claim Adjudication Across the Decision Chain
SummaryThis paper introduces InsClaimBench, an end-to-end benchmark for testing large language models on insurance claim adjudication across auto, property, and health insurance. It contains 3,780 cases and evaluates performance from atomic rule judgments through adjudication modules to payout decisions and amounts, including controlled factual changes. Evaluation of six LLMs shows reliability declines along the decision chain, with joint decision-amount accuracy of 47.54–73.15% versus 74.23–80.19% for payout decisions alone.
In briefQuasi-Binarized Autoencoders: A Bottleneck for Medical Image Anomaly Detection
SummaryThis paper introduces a quasi-binarizing layer that squashes latent values into [0, 1] and adds Laplace noise, bounding the information passed between encoder and decoder. The resulting QBAE model uses one architecture and configuration across seven MedIAnomaly datasets, reaching a mean image-level AUROC of 0.828. It reports the best published results on BraTS2021, with AUROC 0.911 and pixel-level AP 0.838.
WhyIntroduces a quasi-binarizing layer with a provable mutual-information bound, validated on seven datasets with strong AUROC gains over reconstruction-based baselines.
Shouhei Hanaoka, Takahiro Nakao, Atsushi Takamatsu et al.
In briefShared or structured inputs can make reasoning AI agents' collective random choices biased.
SummaryThe paper tests six reasoning models for threshold and divisibility rules in identifier-based random choices. Shared identifiers produced correlated participation among agents, while distinct identifiers with common timestamp bits produced biased participation. Explicitly instructing independent randomization reduced but did not eliminate the shared-input correlation. The authors argue that selection rates alone miss these vulnerabilities, so input-dependent bias, correlation and predictability should be central targets in agent evaluation.
In briefSAPD: A Rollout-Free Method for Post-Training Language Models Using Step-Aligned Supervision
SummaryThe paper proposes SAPD, a rollout-free self-distillation method that turns fixed reference solutions into step-aligned distributional supervision for reasoning. On mathematical reasoning benchmarks it outperforms supervised fine-tuning and label smoothing on average and stays competitive with on-policy reinforcement learning and self-distillation. The authors also report largely preserved out-of-domain coding performance and roughly 2x training-loop speedups over on-policy baselines, and release their code.
In briefDeep learning applied to large-scale seismic interference removal in Angola's Camie field
SummaryIn marine seismic surveys, energy from nearby external sources can contaminate recordings as coherent noise called seismic interference (SI). This paper presents a real-world case study applying a previously proposed deep neural network workflow to attenuate SI across a survey block of over 345 km² in Angola's Camie field. The authors compare the DNN workflow with a conventional geophysical algorithm across the whole block, assessing processing quality and time. The results report higher SI removal accuracy with less signal leakage and more complete removal. The authors also discuss the study's limitations.
In briefAlice: A Large-Scale German Benchmark for Rubric-Based Multi-Dimensional Short Answer Scoring
SummaryThis paper introduces Alice, a large-scale German dataset for automatic short answer scoring based on rubrics, with three subtasks covering learning performance, knowledge elements, and skills. The authors frame rubric-based scoring as a rubric-retrieval task and benchmark models ranging from encoder-only models to lightweight LLMs, plus zero-shot LLM prompting. Results show that zero-shot LLMs struggle to score knowledge elements and skills, and that rubric text is often helpful, especially for those two subtasks.
Zhifan Sun, Sebastian Gombert, Jannik Lossjew et al.
In briefGrading short answers by treating rubric spans as label representations
SummaryThe paper proposes RUSPAN, a framework for automatic short answer scoring that treats rubric descriptions as semantic label representations. It encodes the question, student answer, and all candidate rubric levels in one sequence and scores the levels listwise from a single language model pass. A variant, RUSPAN-RIM, uses a Rubric-Independent Mask so rubric spans do not attend to each other, aiming at zero-shot transfer. On six benchmarks in English, German, and Portuguese, RUSPAN improves over baselines, and RIM with position reindexing gives consistent gains on the held-out PT-ASAG benchmark.
Zhifan Sun, Sebastian Gombert, Fabian Zehner et al.
In briefMeshSIPP: Fast Path Planning for Moving Obstacles
SummaryThis paper presents MeshSIPP, a planner that computes spatiotemporal paths for autonomous navigation among predictable moving obstacles while respecting non-holonomic motion constraints. It groups motion primitives that sweep the same regions, screens them with lightweight checks, and delays exact departure-time search until needed, with time-aware pruning removing redundant branches. The authors prove the search is complete and optimal. Experiments on over 6,000 benchmark instances and ROS 2 simulations report up to a 3x speedup over state-of-the-art spatiotemporal planners.
In briefWhen Rank Rises as LLMs Degrade: Why Spectral Health Monitors Can Mislead
SummaryThis paper examines whether spectral statistics such as RankMe can monitor the health of language model representations during post-training, since rank is often assumed to fall as representations degrade. In controlled experiments on Qwen3-0.6B, data duplication degraded held-out loss while increasing RankMe, so the assumption fails. The authors also test a two-sided, multichannel sequential monitor that identifies damage regimes but does not warn earlier than held-out loss.
In briefAutomatically Building and Updating a Knowledge Graph of MLIP Models
SummaryThis paper presents an LLM-based process that automatically builds a knowledge graph for machine learning interatomic potentials (MLIP), which are used to predict material properties. The process extracts information from documents and articles, then uses a validation loop with SHACL constraints to detect and correct errors, handling each model separately to keep representations consistent. The authors illustrate the approach with a graph built from models listed on the Matbench Discovery leaderboard and show several queries that can be run on it.
In briefCircuitATLAS: Agent-based reasoning over a neuroscience knowledge graph to find drug targets for circuit disorders
SummaryThis paper presents CircuitATLAS, a knowledge graph of systems neuroscience with provenance tracking, paired with an agentic framework for finding drug targets in circuitopathies. The agent reasons from measurable disease phenotypes through neural circuits and cell types to molecular interventions. The authors report that expressing ATP1A3 in interneurons abolished beta- and gamma-band responses to a focal 4-aminopyridine challenge in vivo.
In briefOn-policy distillation teaches models new reasoning skills, not new facts
SummaryThis paper examines whether on-policy distillation (OPD) lets language models acquire new factual knowledge or multi-step reasoning skills. Using a controlled synthetic framework, the authors separate these two capabilities across four models and also test factual QA and competition mathematics. Reverse-KL OPD transfers compositional reasoning skill to unseen structures but transfers little factual knowledge, while switching to forward KL restores factual transfer. The authors conclude that OPD does not expand a model's parametric knowledge but improves how it organizes and composes existing knowledge.
WhyDisentangles factual knowledge from compositional skill transfer via controlled synthetic tasks, revealing reverse-versus-forward KL asymmetry, validated across four models and benchmarks.
In briefTracing Inputs to Verify Who Shaped AI Music
SummaryThis paper examines how to verify which audio sources contributed to an AI-generated music output and whether they influenced it. The authors condition generation solely on audio, trace each output back to its inputs, and measure their musical effect. In their tests, the MixAudio generator's stems follow the prompt audio in timbre and the context audio in harmony. A memorization audit using the musicDNA version identification model found few reproductions outside the recorded inputs.
In briefCoding-agent benchmarks should mirror real users' task flows
SummaryThe authors analyze 4,782 real agent sessions from JetBrains IDEs and find that long interactive sessions mix many task types and switch between them, unlike issue-derived benchmark tasks. They propose SWE-TaskFlow, a method that reshapes existing benchmarks toward a target task flow while keeping verified tasks and tests. In a pilot on 700 SWE-Bench Pro tasks, solving tasks in several sequential steps roughly doubled agent cost without a stable change in resolve rate.
In briefHow Agentic LLMs Decide to Call Tools: A Suppression-Shaped Vector
SummaryThis paper studies the internal mechanism that determines whether an agentic LLM calls a tool or responds directly. The authors convert long prompts into minimal pairs where changing one verb flips the decision. They identify a vector that is causally necessary and sufficient, and find that analysis verbs suppress the tool-call prior. The mechanism recurs across seven models from three families.
WhyMinimal contrastive pairs isolate a causally necessary and sufficient tool-call vector, revealing suppression, replicated across seven models.
In briefDoes training history predict a network's future learning better than its current state?
SummaryThe paper asks whether a neural network's training history predicts its future learning better than its current state. In the main study, small multilayer perceptrons trained under three history regimes were tested with a short probe, and a history state did not improve on a calibrated current-state model. A companion screen on 1,560 synthetic regression runs found history models outperformed current validation error early in training but not later.
In briefIdentifiability and Observability of Deep Normalized Attention: which parameters a deep attention layer's input-output function can uniquely determine.
SummaryThe paper studies which parameters of deep, unmasked, single-head attention are determined by its input-output function. For known positive nonconstant real-analytic normalizers, the function generically determines effective scores and the combined value map, up to signs from even normalizers, proving the real-analytic case of a Henry–Marchetti–Kohn conjecture that includes softmax. It also classifies exceptional fibers and gives sharp Taylor orders for local identification.
In briefTestJack: Can coding benchmark results be trusted? Auditing agentic coding benchmarks through evaluator evolution
SummaryThe paper argues that most coding benchmarks judge solutions only by fixed unit tests, which can miss required behavior and let agents pass by gaming the evaluator. The authors propose TestJack, which generates requirement-targeted tests for each trial, keeps only tests the ground-truth patch passes, and re-examines failures with replayable tests. Across 6 frontier model backends and 5 benchmarks, about 34.4% of trials judged correct violate task requirements, lowering the resolution rate from 50.6% to 33.2%.
In briefSecond-order optimization for variable projection SVMs and road anomaly detection
SummaryThis paper proposes a new second-order optimization framework for minimizing variable projection functionals. It shows the framework is especially useful for training variable projection based kernel methods, such as variable projection support vector machines (VP-SVMs). The authors apply second-order trust region algorithms to train VP-SVM models that recognize road surface abnormalities from 1D signals collected by a tire sensor.
In briefAuditing duplicate subject identities in national brain MRI databases
SummaryNational MRI repositories combine data from many studies, but the same person can appear under different subject identifiers, creating identity duplication. The paper presents HAPPEN, a human-in-the-loop pipeline that combines SHA-256 exact-duplicate detection with supervised contrastive embeddings to retrieve near-duplicate candidates for human review. Deployed on 95,129 scans, it found 1,316 exact-duplicate groups and 1,275 near-duplicate groups, and recovered all 54 genetic-reference pairs. An independent institution also completed the workflow locally on 22,386 scans without retraining.
Jiheng Li, Michael E. Kim, Trent M. Schwartz et al.
In briefSequential Pretraining Favors Large Models: Exposure Therapy Helps Small Models Learn Late Data
SummaryThis paper asks why large neural networks acquire capabilities that small models fail to learn. The authors define and measure "primacy bias," the extent to which exposure to early data distributions impairs later learning. They find that small models allocate learning capacity inefficiently toward early distributions, while sufficiently overparameterized models are robust to this effect. They introduce Exposure Therapy, a simple regularization method, which the authors report improves performance on late data distributions for models up to the billion-parameter scale.
In briefDecoupled optimization for teacher-student semi-supervised learning using a Pioneer Student
SummaryThis paper identifies two optimization problems in semi-supervised learning: tight coupling between teacher and student parameters, and gradient imbalance between labeled and unlabeled losses that pushes shared parameters toward labeled-dominated local minima. The authors propose the Pioneer Student (PiS), an auxiliary branch that works in an independent parameter space and periodically transfers accumulated knowledge back to the teacher-student model. Experiments reportedly show PiS is a plug-and-play module that consistently improves mainstream SSL methods.
In briefWhich Language Should a Skeleton Speak? Choosing Languages for Multilingual Reasoning
SummaryThis paper studies how the language used in skeleton-based reasoning prompts, a training-free way to structure large language model reasoning, affects multilingual mathematical reasoning. The authors propose LASEF, a language-aware skeleton exploration framework, and test it across math benchmarks, model scales, and languages. English skeletons show a small average advantage, most visible for smaller models and low-resource languages, but few language-level gains remain significant after correction and English is not universally best. The effects depend on context, so skeleton language is a design variable requiring multi-level exploration.
In briefA lightweight, general-purpose learned optimizer that reuses averaged gradient history
SummaryThis paper proposes a lightweight, versatile learned optimizer that dynamically recombines gradient history, stored as averages over disjoint time spans. It reduces the prediction space to one scalar coefficient per gradient average, shared across parameters, and progressively averages older gradients to limit memory cost. The authors report zero-shot generalization to unseen tasks and test accuracy gains over Adam on several models, with low FLOPs overhead.
In briefSafeEvo: How Language Models Learn Safety and How It Evolves
SummarySafeEvo is an interpretability framework that studies safety mechanisms in language models at the level of circuits, which are sparse subgraphs inside the model. It identifies weak refusal circuits in pretrained base models, and removing them eliminates the model's refusal of harmful inputs. The authors also propose Safety Circuit Alignment (SCA), which limits safety updates to these circuits, and report across three models and two alignment algorithms that it improves alignment, reduces over-refusal, and retains most original capabilities.
WhyReveals refusal circuits in pretrained models and their evolution during alignment, proposing refusal-only updates; across three models, harmfulness drops 63.21% while utility retains 99.58%.
In briefSTORK: A neural network that observes uterine contractions across space and time in fetal MRI
SummaryThis paper addresses detecting uterine contractions in fetal MRI, which are usually identified manually and discarded. The authors propose STORK, a weakly-supervised multi-instance learning model trained only on coarse series-level labels rather than frame-level annotations. On around 700 multi-vendor dynamic fetal MRI series, it achieves a series-level AUROC of 95.0% and AUPRC of 94.6%, outperforming 3D ResNet and ConvNeXt baselines.
Melissa Schween, Tristan Gottwald, Jordina Aviles Verdera et al.
In briefCOPC: a coupled correction method for stale-data training of language models
SummaryAsynchronous reinforcement learning speeds up post-training of large language models by separating rollout generation from optimization, but it trains on outdated trajectories. The paper shows that correcting only the policy side is insufficient, because advantage estimates also inherit mismatch from behavior-policy continuations. The authors propose Coupled Off-Policy Correction (COPC), which coordinates policy-side and advantage-side correction and reports the highest results on tool-integrated math reasoning and search tasks, with improved training stability.
In briefComparing structured pre-generation elicitation with single-shot prompting for AI-assisted enterprise decisions in a randomised experiment
SummaryThe study compares an interactive scaffold that asks users to clarify context, choose a strategic direction and explain reasoning before AI generates a strategy document, against single-shot prompting. With the scaffold, mean composite quality was 32% higher and off-task deliverables fell from 34% to 5%, at about 2.4 minutes extra time. The authors note that delayed retention, error detection and effects in live organisations remain to be studied.
In briefScalable Bayesian density regression using kinetic Langevin sampling
SummaryThis paper proposes a scalable Bayesian method for conditional density estimation, which models the full distribution of a response given covariates, such as photometric redshifts of galaxies. It uses a logistic Gaussian process for the log conditional density and samples the latent field with kinetic Langevin dynamics rather than Laplace or variational approximations. The method trains on a single GPU and is evaluated on photometric-redshift benchmarks with up to 3.9 million training observations. On density and calibration metrics, it is competitive with state-of-the-art tabular foundation models.
Daniel Paulin, \'Ad\'am Jung, and Andr\'as A. Bencz\'ur
In briefLearning when LLM agents should think: situation-conditioned thinking policies for long-running agents
SummaryThe paper proposes a memory framework that turns past reasoning experience into a lightweight policy predicting what should be thought about in the current situation, leaving detailed reasoning to an LLM. Temporary experiences are analyzed across multiple episodes to find repeated long-range regularities, which are consolidated into new thinking knowledge. Reported experiments include 1.000 F1 on temporal-rule generalization, DeepSeek reasoning F1 rising from 0.789 to 0.868, and per-query processing time falling from 0.3636 ms to 0.0382 ms at 30,000 historical situations.
In briefTwo AI Suggestions Versus One: A Randomized Study on Resident X-ray Reading
SummaryThis study compared whether residents reading radiographs did better with one AI suggestion or two when the AI was sometimes wrong. In a randomized multicenter trial, 123 residents with under 3 years of experience read 60 X-rays under three support conditions. Radiology residents improved more with dual suggestions, and when GPT-5.4 erred, dual-suggestion accuracy was higher than single-suggestion accuracy in both radiology and non-radiology residents. The authors suggest dual suggestions may reduce the influence of incorrect AI advice, with the benefit seen mainly in radiology.
In briefAdaptive Code Generation for Robot Control: LLMs Turn Natural-Language Goals into Verified Robot Programs
SummaryThe paper presents an architecture in which a large language model translates high-level natural-language intentions into executable program code limited to a formal robot library and checked for valid syntax. A vision-language model supplies semantic grounding, and replanning is triggered by geometric and semantic thresholds alongside continuous runtime monitoring. The authors report benchmarks across frontier models, claiming that this constrained, reactive loop enables robust completion of complex goals in dynamic, unknown environments.
Justus Flerlage, Thorsten Wittkopp, Alexander Acker et al.
In briefTeachers critique each other to distill reasoning into small models
SummaryThis paper presents Collaborative Reasoning Distillation (CRD), a framework in which teacher models critique each other's reasoning and step-wise quality is assessed independently of final answers. The coherent steps are then combined to train compact student models. The authors report that their 4B model, CRD-4B, scores 97.3% on MATH-500 and 70.3% on AIME'25 using 50K training examples.
In briefJevForest: Path Voting for Choosing Which Features to Observe Under a Budget
SummaryThe paper proposes JevForest, a policy for deciding which features to acquire when observation budgets are limited. It combines path-dependent proposals from bootstrapped trees, weights them by global information gain, and predicts with a shared masked classifier. Small experiments on AG News, TREC, and MiniBooNE show results that vary by task, and the authors state these findings do not establish a general advantage for path voting.
In briefGradient-based optimisation of continuous scanner poses for sparse-view cone-beam CT
SummaryThis paper proposes a gradient-based method for planning cone-beam CT scan trajectories, treating each X-ray source pose as a continuous variable moved jointly on the scanner's kinematic manifold. The objective combines soft-Tuy plane coverage, a continuous view covariance loss, and an analytic attenuation-aware ray-bundle penalty. On a Defrise flange phantom, the authors report recovering laminar defects invisible to a circular orbit, and the method works across circular, C-arm, two-axis and free-sphere scanner geometries. A robot CT bench experiment showed feasibility but no consistent metric gain over uniform band sampling.
Linda-Sophie Schneider, Simon Wittl, Gabriel Herl et al.
In briefCompressing visual tokens for faster document retrieval using optimal transport
SummaryVision-language models can search visual documents without OCR, but they produce too many tokens, which limits scalability. This paper proposes a visual token compression framework based on Unbalanced Optimal Transport that jointly handles token importance, semantic redundancy, and coverage in one optimization process. The authors show theoretically that minimizing transport cost upper-bounds the MaxSim retrieval error, and experiments on ViDoRe V1 and V3 report better efficiency-effectiveness trade-offs than baselines.
In briefOnline resource allocation with a state that depends on decisions; fewer optimization solves can perform better
SummaryThis paper studies finite-horizon online resource allocation where requests are i.i.d. and actions influence a finite Markov state that affects future rewards and resource use. The authors analyze regret for infrequent and frequent re-solving under known and unknown request priors, finding that more frequent re-solving can perform asymptotically worse when the optimum is degenerate. They propose a three-phase U-shaped policy using only O(log log T) LP solves. Numerical experiments support that infrequent schemes are generally more robust.
WhyProposes a new framework for resource allocation with endogenous Markov states, proves regret gaps under degeneracy, and gives an O(log log T)-solve U-shaped policy with strong theory.
In briefPEACE: Learning nonadiabatic molecular dynamics with symmetry-aware electronic representations
SummaryThis paper introduces PEACE, a method that combines a parity-equivariant latent Hamiltonian with a learned electronic connection to model nonadiabatic molecular dynamics. Ablation studies show that symmetry-allowed state mixing and electronic-frame variation each contribute to reproducing crossing structures and relaxation dynamics. PEACE closely matches excited-state population dynamics from first-principles simulations, and its extension to spin-orbit coupling enables simulation of intersystem crossing.
In briefHow Do LLMs Change Predictions Under Negation? A paper on why language models often repeat the original answer when asked what something is not, and a training fix.
SummaryThe authors evaluate recent open- and closed-source LLMs on a negation benchmark and find that in 37-71% of cases the models repeat the original answer under negation. A mechanistic analysis shows that specialized attention heads and MLP neurons implement negation by suppressing the original answer while promoting a favored alternative in the same category, unlike human processing that uses the original answer to determine exclusions. The authors propose a training objective requiring larger shifts in answer preference for more confident original predictions, which they report reduces negation failures with less loss of general capability than standard fine-tuning baselines.
WhyIdentifies a suppress-and-promote negation mechanism, then proposes a training objective, supported by benchmark evaluation and mechanistic analysis.
In briefRELATE: A framework for measuring whether LLMs steer users toward or away from real-world relationships
SummaryThis paper examines whether large language models used for emotional support encourage users to rely on the AI itself or to connect with people in their lives. It defines relational orientation through inward-facing language, which positions the AI as the ongoing source of support, and outward-scaffolding language, which encourages human connection. The authors introduce RELATE, which uses 228 stimuli to evaluate seven LLMs over multi-turn dialogues. Results show inward-facing sentences are more frequent at the sixth assistant turn than the first, and outward-scaffolding sentences are much lower for hesitant, indirect simulated users.
In briefEvoSignal: LLM-guided evolutionary design of modular traffic signal control programs
SummaryThe paper treats traffic signal control as a modular program design problem and proposes EvoSignal, an evolutionary framework guided by large language models that uses traffic knowledge and performance feedback to improve control programs. The resulting programs run without online LLM inference. In simulations across five scenarios on two real-world road networks, the selected default program reduced waiting time by 16.8% to 49.2% relative to the lowest waiting time among 20 baselines in each scenario.
In briefWorld Potential Model: Using Pretrained World Knowledge to Measure Task Progress
SummaryThe paper addresses long-horizon language agents that receive supervision only from final task outcomes, proposing a World Potential Model (WPM) that uses pretrained models' existing world knowledge to judge task progress. In ALFWorld and ScienceWorld, off-the-shelf pretrained models without task-specific fine-tuning substantially outperform chance at recovering progress structure. Anchoring these judgments to task milestones yields scalar world potentials whose temporal differences give step-level credit for policy optimization, and WPM-guided optimization improves success over outcome-only GRPO in all evaluated configurations.
In briefUniCSI: A Universal Wi-Fi CSI Encoder for Everyday Human Sensing
SummaryThis paper presents UniCSI, a unified architecture that processes Wi-Fi channel state information (CSI) directly, even when it varies in subcarrier count, bandwidth, and carrier band. It uses a physics-informed RF tokenizer that encodes frequency channels by their position in the physical spectrum, and a spectral aggregator that maps variable-length channel sequences to a fixed-size representation. Evaluated on 25 heterogeneous public datasets, the approach reportedly improves cross-domain transfer under both supervised and self-supervised training.
In briefDrugTargetWorld: A Synthetic Biobank for Training and Evaluating AI Scientists
SummaryThis paper presents DrugTargetWorld, a framework that procedurally generates simulated biobanks with known but hidden causal structure for drug target discovery. Each simulated world includes genetic, protein, health record, outcome, and synthetic MRI data for 54,000 participants, and agents must identify causal driver proteins and their beneficial modulation direction. In evaluations across 540 episodes on 20 cardiovascular worlds, the top two agents recovered 64% of causal drivers on average, but no agent reliably distinguished misleading non-causal proteins.
Samuel Margolis, Paul Schmiedmayer, Alan Huang et al.
In briefReliability of LLM Judges for Evaluating Entity Alignment
SummaryThis paper examines whether large language models can reliably act as judges for evaluating entity alignment, the task of matching equivalent entities across knowledge graphs. Testing three frontier models, three datasets, and four alignment systems, the authors find an anchor bias: when the system's decision label is visible, judge discrimination drops sharply. A label-free protocol largely recovers performance, and the authors release a biomedical benchmark (MeSH-SNOMED CT, 15K pairs) and an auditing framework.
WhyFirst systematic study of LLM judge reliability for entity alignment, identifying anchor bias with counterfactual and human-annotation evidence; valuable for the subfield.
In briefA framework for systematically reviewing machine learning assets in AI registries
SummaryThis paper explores how systematic review methods from scientific literature can be applied to retrieving machine learning assets such as pre-trained models, datasets, and benchmarks. The authors adapt the planning, conducting, and documenting phases to AI registries and propose a framework with registry-aware search strategies, cross-registry schema alignment, and dependency-driven asset exploration. They argue that ML asset retrieval should be a systematic and reproducible process rather than an ad hoc activity.
Alexandra Gonz\'alez, Quim Motger, Xavier Franch et al.
In briefCircuitGate: Circuit-level functional modeling of AIGs that stays logically consistent
SummaryThis paper introduces CircuitGate, a learning framework that represents And-Inverter Graphs (AIGs), a standard structure used in chip logic synthesis and verification. It moves beyond local gate-level message passing by encoding global primary-input support and modeling overlapping support at reconvergent fanins, along with Boolean constraints. The authors report that it outperforms existing methods on equivalent-gate identification and signal-probability prediction, with MAE reductions of up to 21.7% and 14.2%.
In briefVerification and Self-Improvement in Agentic AI: Foundations and Limits
SummaryThis paper examines how agentic AI systems improve through longer search, added support, or changed verification methods, using a bounded verification framework with hidden random bits. It proves that independent majority amplification preserves the defined languages, while existential acceptance over random tapes can admit incorrect outputs. It also gives complexity class inclusions and discusses limits for recursive self-improvement.
In briefAbstention-Based Certification: Distribution-Free Guarantees for Chain-of-Thought Verifiers
SummaryThis paper examines what distribution-free selective guarantees can offer for verifiers that predict whether chain-of-thought reasoning is correct, when only tens to a few hundred labelled problems are available for calibration. Using seven open models, five verifier signals and 37,000 graded traces, the authors find that standard certificates often abstain, and when they do fire, their failure probability can be much higher than the nominal bound. They also propose a floor-started fixed-sequence certificate that requires no monotonicity assumptions and gives higher coverage than the Bonferroni baseline.
In briefWAPR: A Foundation Model for Wide-Angle Pose Refinement of Unseen Objects
SummaryThis paper presents WAPR, a zero-shot model that refines candidate 6D object poses with rotational errors of up to 90 degrees. It uses rotational symmetry priors to canonicalize symmetry-equivalent pose targets and introduces SA6D, a training dataset of about 50K augmented object instances and about 2M rendered RGB-D images. Experiments on seven BOP core datasets report state-of-the-art performance for unseen-object 6D pose localization and detection.
WhyProposes wide-angle zero-shot pose refinement with symmetry priors, validated on seven BOP datasets with state-of-the-art results; useful for the subfield.
In briefKASALv2: Fully Automatic Classification and Localization of 3D Rotational Symmetry Axes
SummaryThis paper proposes a fully automatic, reference-free framework that classifies symmetry types, identifies rotational orders, and localizes full symmetry axes for 3D objects across all eight canonical rotational symmetry types. It achieves 94.75% accuracy on 438 symmetric objects in GSO and performs strongly on idealized and real-world datasets. Using these automatically estimated symmetry priors to train FoundationPose improves accuracy by up to 0.9% across five BOP datasets.
WhyProposes automatic reference-free rotational symmetry classification and axis localization, validated on GSO and improving FoundationPose across BOP datasets.
In briefUnpaired Canonical Correlation Analysis: learning maximally correlated projections without any paired samples
SummaryCanonical Correlation Analysis (CCA) is a standard method for learning shared spaces across multiple views, but it requires paired data, which is often unavailable. This paper proposes Unpaired CCA (UCCA), which learns linear projections that maximize correlation with the true underlying pairing without using any paired samples in training. The authors connect the Quadratic Assignment Problem (QAP) to CCA theoretically and use this to derive a practical method that works only from unpaired data. Experiments on real-world multi-modal datasets show UCCA significantly outperforms recent unpaired alignment baselines at recovering the true correlation.
In briefCode understanding is a bottleneck for coding agents: a study finding that reading and analyzing code, not just editing it, is what makes tasks hard for AI coding tools.
SummaryThe paper introduces CABRA, a synthetic benchmark that builds coding tasks as call graph transformations and scales difficulty across four axes. Across 6,840 tasks, eight LLMs and six coding agents were tested, showing that LLM accuracy drops as tasks grow while agents stay near-perfect by offloading work to tools. On SWE-bench Verified, tool call counts predicted agent accuracy better than lines of code edited, suggesting code understanding is the harder part.
In briefComparing Evaluation Metrics for Summarizing Long Financial Reports with Transformers
SummaryUK annual reports from London Stock Exchange firms average about 80 pages, and this study benchmarks pre-trained transformer models and extraction methods on the FNS 2020 shared-task dataset. It examines how different evaluation metrics behave and whether they reflect true summarization ability. The authors propose BRUGEscore, the harmonic mean of ROUGE-2 and BERTscore, and check their results with a statistical significance test and an adversarial analysis using three corruption methods.
In briefCalibrate Before Fusing: Reference-Space Alignment for Supervision Without Labels
SummaryThe paper proposes a calibration-first framework that generates supervision scores without ground-truth labels or a shared annotation space. It aligns subset-specific scorers to a synthetic ordinal reference space, built from ordered calibration features, before fusing them. Across three benchmark datasets, the framework consistently outperforms uncalibrated averaging and achieves higher primary-metric point estimates than the best individual scorer. Results relative to sample-dependent baselines vary by domain.
In briefReflected Anchored Langevin Algorithms for sampling non-differentiable targets on constrained domains
SummaryThis paper introduces reflected anchored Langevin dynamics (RALD), a reflected diffusion that converges to non-differentiable target distributions on constrained domains. It uses a smooth anchored reference potential and multiplies the drift and noise covariance by the same state-dependent scaling factor. Euler-Maruyama discretization with projection yields the RALMC algorithm. The authors prove explicit convergence bounds and iteration complexity in the 2-Wasserstein distance, and numerical experiments illustrate the theoretical predictions.
In briefNot All Uncertainty Matters: Simulation-Based Fast-Slow Reasoning for Autonomous Driving
SummaryThis paper presents SIGMA, a framework that uses simulation to test how uncertain scene elements affect feasible trajectories and planning cost, deciding when cloud vision-language models should be consulted. It introduces expected planning gain (EPG) as a decision metric for cloud invocation and request prioritization. In CARLA experiments, compared with fixed-period collaboration, SIGMA reduces unnecessary cloud interactions by 50%, improves navigation success by more than 6%, and cuts finish time by up to 26.2% in dynamic scenarios.
In briefDifferential refresh of derived data: optimal per-partition allocation, validated on materialized views
SummaryThe paper argues that refreshing derived state such as materialized views, cached aggregates, or trained models using a single weighted staleness threshold on a globally recomputed artifact is equivalent to a periodic timer and cannot outperform it. It proposes refreshing partitions, such as GROUP BY groups, per-tenant views, or data segments, independently, with the optimal per-partition recompute rate proportional to the square root of partition risk. Experiments on SQLite with real SQL show a 14 to 15 percent reduction in realized weighted staleness at matched recompute counts.
WhyShows global staleness triggers equal uniform timers, derives square-root-of-risk optimal per-partition refresh, and validates a 14-15% staleness reduction on SQLite.
In briefActiveLang: Robots Mapping 3D Scenes with Language Labels via Uncertainty-Guided Exploration
SummaryThis paper presents ActiveLang, an autonomous system that lets robots build 3D maps annotated with language while exploring unfamiliar environments. It adapts language features online on a compact dual-Gaussian representation, jointly reconstructing geometry, appearance, and open-vocabulary semantics with modest memory. Its planner selects informative viewpoints to map effectively with fewer observations and lower computation. Experiments on Replica and ScanNet++ show improved 2D and 3D open-vocabulary segmentation over online and offline baselines.
In briefNot Seeing the Hazard: Frozen VLM Safety Score Tracks Its Captions
SummaryThe paper evaluates a frozen CLIP prompt-margin safety score used as a reinforcement learning signal for detecting hazards. The authors apply it to trajectories from policies that never receive the score and compare pre-contact observations with contact-free ones, while varying captions, encoder, and camera view. They report that the score mainly reflects resemblance to the scene shared by its captions. Because a constant-confidence control retains the lower catastrophe-rate estimate, policy gains do not establish hazard perception.
In briefSTRIKE: learning how a scene's visual state changes for physical world modeling
SummaryThis paper presents STRIKE, a framework that separates learning visual state transitions from dense video generation. An image-based transition model predicts the next scene state from the current image, a transition description, and elapsed time, and is applied recursively to produce future visual states. A separately trained dynamic model then generates the full video conditioned on these states. Experiments on Physics-IQ Verified, PhyGenBench, Pisa-Experiments, and RoboTwin2.0 report gains over video-backbone baselines in physical consistency and manipulation-video fidelity.
In briefOmniCam: Generating Camera Paths from a Panorama and Text via Geometry-Aware Learning
SummaryThis paper introduces OmniCam, an autoregressive model that generates camera pose sequences from a single panorama and text descriptions of trajectories. It combines a panoramic point-cloud encoder, hybrid rotation-and-translation tokenization, and separate geometric and semantic conditioning with a 3D target anchor. The authors also release OmniCaT, a dataset of 267,700 trajectories across four camera behaviors. On their reported evaluation, OmniCam reduces trajectory error and collision rate relative to retrained baselines.
In briefLiG-DETR: A Method That Reassembles Local Features into Global Representations for Aerial Object Detection
SummaryAerial images contain objects of widely varying scale and density, and small objects are easily degraded by downsampling and feature compression. The paper proposes LiG-DETR, which acquires high-fidelity local features before degradation, reassembles them by their original spatial locations into a globally aligned local feature level, and jointly decodes global and local features with a single DETR decoder. It also introduces Context-Preserved Selective Reassembly and Density-Aware Adaptive Query Allocation to reduce redundant computation. Experiments report substantial gains on small and medium objects while large-object performance remains strong.
In briefPhysics-Informed Neural Plasticity: PDE Solvers That Reshape Their Own Structure
SummaryThis paper introduces physics-informed neural plasticity, in which a solver's representation reshapes during optimization in response to unresolved physics. The authors instantiate this idea as ReCAP, a Gaussian-localized PDE solver that redistributes capacity through local enrichment, residual-directed splitting, gate-based pruning, and function-aware merging. On five challenging 3D and 4D PDE benchmarks against 11 physics-informed solvers, ReCAP achieves the lowest relative L2 error on every problem, reducing error by 10.7% to 27.5% relative to the strongest competing result.
Chun-Wun Cheng, Bingcheng Hu, Angelica I. Aviles-Rivero
In briefTERRA: Learning Transferable Latent Robot Actions via Temporal Effect Representation
SummaryThis paper studies latent actions, action-like codes inferred from visual transitions, used to train robot policies. The authors propose TERRA, which describes a transition by a compact temporal effect combining net feature change and low-order within-window dynamics, and learns a continuous latent from it. They also introduce Effect-Anchored Transport (EAT), which decodes latents in other initial states and anchors their effects to the source effect. With frozen linear readers, the paper reports TERRA predicts actions more accurately than UniVLA and a LAPA-style baseline, and reaches 93.4% average success on LIBERO versus 91.8% for UniVLA.
In briefSafe on average but risky in the worst cases: can episodic-cost tails be controlled?
SummaryThis paper studies tail risk in safe reinforcement learning, where policies are usually constrained only on mean episodic cost. The authors measure the worst 10% of episodes with CVaR at 0.1 and classify policies as tail-safe when this value is within budget. They evaluate five standard algorithms on three Safety-Gymnasium navigation tasks to identify tail-unsafe policies, then examine whether these violations can be controlled while preserving return.
In briefBias-aware calibration and optimal control for stochastic bioprocess digital twins
SummaryThis paper presents a framework for calibrating and controlling digital twins of multiscale bioprocess models, represented as stochastic differential equations and fitted to sparse observations. It uses moment expansions to characterize parameter bias and adjoint methods to propagate calibration uncertainty to policy performance. An Actor-Simulator algorithm jointly updates parameters, selects informative experiments, and optimizes control policies. Numerical studies report improved calibration accuracy, sample efficiency, and control performance over state-of-the-art baselines.
In briefRT-DETR-World: Transferring LLM Semantics to Real-Time Open-Vocabulary Detection
SummaryThis paper proposes RT-DETR-World, a compact DETR-style detector that recognizes object categories unseen during training while running in real time. During training, it uses category names, object descriptions, and image descriptions as supervision, and a Dual-Path Description Alignment method transfers semantics from an LLM teacher into the lightweight model. The teacher modules are removed at inference, which keeps only lightweight query-text matching, and experiments report competitive zero-shot accuracy with a favorable accuracy-efficiency trade-off.
In briefSpatialUQ: Post-Hoc Uncertainty Estimation for Black-Box Vision Models Using Spatial Consistency
SummaryThis paper introduces SpatialUQ, a method that estimates uncertainty in frozen black-box vision models using only output probabilities. It measures the divergence between a global prediction and the average prediction over five fixed spatial crops. On the NIH ChestX-ray14 dataset, the proposed Multicrop Uncertainty Score achieves higher failure-detection AUC than MC-Dropout at about one-fifth the compute. The authors note the method suits diffuse findings but is less reliable for small focal lesions such as nodules, and its relationship with model quality weakens under severe distribution shift.
WhyProposes a black-box post-hoc uncertainty method via spatial crop consistency, beating MC-Dropout on NIH ChestX-ray14 at lower compute with solid experiments.
Md Kawsher Mahbub, Milon Biswas, Mirza Niaz Morshed et al.
In briefReam: A Literature Review Workspace for Human-Agent Mutual Awareness (Research Paper)
SummaryThis paper presents Ream, a literature review workspace that helps humans and AI agents see each other's activity through structured documents, bidirectional engagement tracking, and localized visualizations. In a study with eighteen researchers, participants used these traces to inspect evidence, steer agents, and communicate via annotations. Shared histories also helped agents build on earlier work.
In briefSparse Feature Policy Unlearning: Reducing State Hallucination in Vision-Language-Action Robot Models
SummaryThis paper studies state hallucination, a failure in which Vision-Language-Action robot models act as if an unrealized robot-object state had been achieved. The authors find this coincides with weaker attention to task-relevant visual regions, and sparse autoencoder analysis shows related sparse features activate during these failures. They propose SOUL, which unlearns policy knowledge tied to hallucinated behaviors while retaining knowledge from successful ones. Experiments across VLA architectures in simulation and real-world settings report fewer hallucinated failures and higher task success.
In briefCognitive schemas and task structure for reusable problem-solving knowledge
SummaryThis paper studies how explicit representations can support reusable cognitive schemas in knowledge-based problem solving. It proposes a structural framework that organizes schemas by the information and relations needed to use them, separating context-dependent relations from more stable structures. The authors distinguish inconsistency, underdetermination, and contextual insufficiency as representation limits, and formally show when reducing a representation preserves task-relevant solution structure.
In briefAttribution Blind Spot: Layer-wise Diagnostics of Source Reliance in Retrieval-Augmented Models
SummaryThe paper studies whether retrieval-augmented language models actually rely on the documents they retrieve when knowledge conflicts arise. It uses a signed projection onto a training-fitted first principal component (PC1) of hidden-state changes and compares it with state-change magnitude. Equal-norm interventions along PC1 shift source preference while better preserving non-target behavior. The central finding is that representations that diagnose a model's source choice need not be the ones that best control it.
In briefGaussian Material Fields for Multi-Energy CT Material Decomposition
SummaryThis paper introduces Gaussian material fields, a 3D representation that shares anisotropic Gaussian primitives across materials while giving each material its own nonnegative coefficients. A differentiable spectral forward model jointly optimizes spatial geometry and material composition from multi-energy projections, with material-aware adaptive density control. Across 15 synthesized cases built from public CT data, it improves average PSNR by 4.03 dB and SSIM by 4.96%, and reduces NRMSE by 33.45% relative to the strongest baseline.
In briefGoldsmith: Gold-Loss-Guided Annotation Definition Optimization with an Agentic Pipeline
SummaryThe paper presents Goldsmith, an agentic pipeline that turns a small set of expert-annotated gold examples into a reusable structured annotation definition. Candidate definitions are run on the gold examples and scored with an executable structured loss, and an LLM editor revises them based on the highest-loss failures, keeping changes only when loss decreases. The authors report improvements over several prompt-optimization baselines and better downstream annotation across span, relation, and event-argument tasks.
In briefJevNexus: Treating Schema Correspondences as Decisions for Schema Matching
SummaryThis paper presents JevNexus, a schema matching method that finds corresponding columns across database tables. It combines typed pairwise decisions with schema and instance evidence, and runs listwise refinement only when the evidence disagrees and the fused margin is small. Across 561 cases, it reports MRR of 0.930 and Hits@1 of 0.909, slightly above Magneto's 0.926 and 0.903, with mean latency reduced from 123.452 to 15.929 seconds. Paired analysis finds no statistically significant difference in accuracy.
In briefReducing Accent-Induced L1 Bias in Self-Supervised Speech Representations for Language Identification
SummaryLanguage identification models often misclassify non-native speech as the speaker's first language because accent is confused with language. The authors show that non-native speech representations lie between the native target-language and native L1 poles, which causes this systematic error. They propose a geometric projection that estimates an L1-bias direction from native speech alone and removes it before the frozen LID head. Across five MMS-LID models and non-native corpora, the method substantially improves identification of L2-accented speech while preserving predictions for native speech.
In briefMIMESIS: Training Interactive Agents with Learned User Simulators
SummaryThis paper introduces MIMESIS, a user simulator trained on human conversations with explicit reasoning supervision and 13 realistic behavioral patterns, offered as a scalable alternative to costly human feedback. The authors report that their 9B model scores 65.7 on SOUL-Index, surpassing the strongest frontier model, and improves behavioral fidelity over the baseline Claude-Opus-5. They also freeze the simulator to train agents via multi-turn reinforcement learning and propose Coached On-Policy Self-Distillation (CSD) for denser token-level supervision.
In briefInstanceBench: A Diagnostic Benchmark for Referential Reasoning in Referring Segmentation
SummaryThis paper introduces InstanceBench, an instance-centered benchmark with 6,194 images, 9,264 target instances, and 25,077 human-verified expressions for referring expression segmentation. It tests whether models keep the same target across different valid grounding paths and separates target-selection errors from mask-generation errors. Evaluating 22 checkpoints from 18 model families, the best model reaches 67.1% mIoU but only 59.6% All@0.7, with target selection identified as the main bottleneck.
In briefCHASE: Using channel-aligned structure to design geometry-aware neural network models
SummaryThe paper proposes CHASE, a method that uses Geometric and Spectral Alignment (GSA), which describes spectral concentration, channel alignment and singular bases in trained networks, to guide practical model design. It covers six applications, including CORA, COEC and CORAM for finetuning, pruning compensation and model merging. Three new methods are introduced: CAGA for converting multi-head attention to grouped-query attention, SAKV for KV-cache compression, and CAPS for structured pruning. Reported experiments indicate improvements over representative baselines.
In briefMORA: Separating Real Anomalies from Drift in Time-Series Detection
SummaryThis paper addresses time-series anomaly detection in non-stationary data, where distribution drift and true anomalies can look alike. It proposes MORA, which reconstructs the same local target from short- and long-term views and uses the reconstruction gap to judge whether broader context explains a local deviation. The framework conservatively corrects the anomaly score and needs no drift labels or online adaptation. Experiments on four benchmarks show robustness to non-stationarity while remaining sensitive to genuine anomalies.
In briefWhen Should an In-Context Learner Expand Its Hypothesis Space?
SummaryThe paper studies when a learning system facing prediction failures should expand to a richer model family rather than repair locally. It introduces a Structural Revision Environment that produces matched failures from noise, exceptions, within-family change and out-of-family structure, with exact Bayesian calculations and a normative solution. The authors find revision is a value boundary rather than an evidence threshold, and Transformers trained in the environment reproduce this boundary from utility alone.
WhyFrames hypothesis-space expansion as a costly sequential decision with exact Bayesian solutions, showing the boundary is a value threshold; supported by exact calculations, trained transformers, and l
In briefBoundary-free contextual biasing for unsegmented languages using depth-adaptive gating and reading-space matching
SummaryThis paper proposes a speech recognition decoder that boosts expected words without needing word boundaries, which Japanese and Chinese lack. It works on frozen public CTC models using a character-level Aho-Corasick automaton, requiring no training and no second pass. Two mechanisms, depth-adaptive gating and reading-space matching, replace word boundaries. On the hard R1 subset of Aishell-1 NE it reports 66.5% recall, above a trained CLAS baseline at 64%, and it transfers to WenetSpeech and another architecture without retuning.
In briefLLM-Enabled Drone Dispatch: A Survey and Classification of System-Level Approaches
SummaryThis survey reviews how large language models (LLMs) are used in unmanned aerial vehicle (UAV) dispatch systems. It treats LLMs as a cross-layer semantic orchestration layer linking human instructions, external solvers, and distributed control modules. The literature is organized into four dispatch paradigms: pipeline, global assignment, decentralized agentic, and divide-and-conquer, with analysis of each paradigm's logic, structure, and possible LLM roles. The paper also discusses scalability, robustness, coordination burden, and verification needs for safety-critical deployments.
In briefBanglaRhet: Benchmarking Models for Detecting Rhetoric and Persuasion in Bangla Political Speech
SummaryThis paper benchmarks transformer and classical models for detecting rhetorical devices and persuasion techniques in Bangla political speech. It uses BanglaRhet, a manually annotated corpus of 30,289 speech segments from public political news, and sets up two single-label classification tasks. BanglaBERT performs best, reaching 65.40% macro-F1 for rhetorical technique detection and 66.46% for persuasion technique detection, beating the strongest classical baseline.
In briefRotation-invariant tangent-angle descriptor and band U-Net for predicting 2D fragment adjacency
SummaryThis paper predicts whether pairs of 2D fragments are adjacent, based on their contours, improving a two-stage method from Beaulac's thesis. It replaces image-window scores with comparisons of tangent-angle profiles, making the approach invariant to fragment rotation, and uses a band U-Net to segment the shared arc and classify the pair together. On the synthetic dataset it matches or beats the original approach, reaching 98% accuracy. Trained only on synthetic data, it achieves an AUC of 0.93 on the PairingNet benchmark.
Guillaume Brouillette, Alain Goupil, Pierre-Olivier Paris\'e et al.
In briefRight number, wrong state? Measuring how LLMs swap in other states' policy values
SummaryThe authors test whether LLMs answering state-specific Medicaid income-eligibility questions return values that belong to another jurisdiction, using fixed wording and varying only the state across 51 jurisdictions. Under a pre-registered protocol, Claude Sonnet 5.5 and GPT-5.6 Sol reproducibly gave another state's current value for 10 and 25 of 153 items. The authors argue that measuring cross-jurisdiction error requires a complete same-state reference set, since many apparent substitutions match the asked state's own values under other conventions or years.
In briefDSReg: a method that provably recovers individual world latent variables without reconstruction
SummaryThe paper introduces DSReg (Dependency-Sparsity Regularization), which recovers individual world latent variables up to signed permutation without reconstruction, a decoder, or labels, under a condition called Structural Diversity. It can be applied post hoc to any linearly identified representation, reusing trained checkpoints. The authors claim this yields the first fully identifiable JEPA that recovers every world latent.
Yujia Zheng, David Klindt, Randall Balestriero et al.
