In briefDo ResNets Route? Sparse Interaction Experts in Residual Networks
SummaryThis paper treats a trained ResNet as a set function over binary residual-branch masks and uses Möbius inversion to exactly split its output into single residual corrections and higher-order interactions. Exhaustive analysis of ImageNet-pretrained ResNet-18 and ResNet-34 shows interaction mass peaks at orders five and ten, respectively, rather than at low orders. The authors conclude that dense ResNets perform an implicit soft routing, where all blocks execute but different inputs rely on different residual interaction experts.
In briefTangent Schrödinger Bridge Matching: learning stochastic transport with mechanistic sensitivities
SummaryThis paper introduces Tangent Schrödinger Bridge Matching (Tangent-SBM), which learns stochastic transport from endpoint observations and mechanistic sensitivities by propagating parameter derivatives alongside trajectories. Across Gaussian, double-well, reaction-diffusion, and stochastic Navier–Stokes systems, it improves sensitivity and finite-change prediction over matched conditional-bridge baselines while keeping endpoint and distributional accuracy comparable. In a viscosity selection test on Navier–Stokes, it reduces tracking error relative to taking no action.
In briefViTok: Better dense features in multi-teacher distillation using PHI-S and masked image modelling
SummaryThis paper studies how to combine several training improvements into one multi-teacher distillation recipe that keeps both global recognition and dense semantic quality. A student model distilled from SigLIP2 and DINOv3-L adds split adaptor heads, masked image modelling, and PHI-S feature balancing. The model slightly surpasses the DINOv3-L teacher on ImageNet-1K kNN classification, and PHI-S restores ADE20K segmentation to match the teacher. The authors also report negative results, including inconsistent gains from ImageNet22K and interference from extra teachers such as SAM3 or HOG.
In briefLUMOS: Tracing How LLMs Turn Training Data into Answers
SummaryThis paper introduces LUMOS, a diagnostic framework that traces knowledge from training-data exposure to model outputs, using OLMo 2 because its training corpus is fully transparent. The authors find that rare facts are internally separable (84%) but expressed behaviorally only 54% of the time, a gap that narrows with scale. Self-reflection on unseen content falls to random-baseline levels (49%), and chain-of-thought prompting does not fix this. The authors argue the training-data axis should be a standard part of LLM knowledge evaluation.
In briefInterpreting at Write Time: A Policy Ablation for Multi-Goal Agent Memory
SummaryLong-running assistants must summarise their history, and what a summary keeps depends on the goals it serves. This paper compares three writing policies: a goal-free summary, one summary covering all goals, and a separate summary per goal, tested across several models and event streams. Summaries written for different goals overlap less with each other. Per-goal summaries perform best on relevance, completeness and accuracy, while the all-goal summary even underperforms a neutral summary written with a smaller budget.
In briefWhen Can We Trust the Matching Principle? Robust Deployment Geometry Under Uncertainty
SummaryThe paper uses a trust ratio, tau = epsilon/gamma (estimation uncertainty over spectral separation), to decide whether to match deployment geometry or spread the penalty. Under a linear-quadratic Matching response, drift between estimated and oracle projector matching scales as tau squared in the Davis-Kahan separation region tau < 1/2. It proposes Confidence-Calibrated Matching (CCM), which is directional when tau is small and progressively isotropic otherwise, with thresholds set by calibration rather than the theorem. Experiments include UCI HAR embeddings, where always-match performs worse than abstaining in every cell.
In briefRevealing Epistemic Uncertainty in MLLMs via Causal-Invariant Masking
SummaryThis paper addresses the need for uncertainty quantification in multimodal large language models, which are prone to hallucinations. It proposes Causal-Invariant Masking, which measures semantic shift between original predictions and those conditioned on a causally focused view, using Semantic Divergence as the core metric for capturing epistemic uncertainty from model limitations. The authors also introduce Expected Embedding Drift, a faster geometric proxy. Experiments report state-of-the-art results on various benchmarks, with EED running nearly 50% faster at comparable performance.
In briefMisinformation without triggers: from factual answers to decisions
SummaryThe paper studies whether false content in language model training data affects downstream decisions, not just direct factual answers. Using a controlled Guess the Capital task and misleading Facebook posts about the 2019–20 Australian bushfires, it compares false training with matched truthful controls. The authors find that correct direct answers do not guarantee correct decisions, and removing the injected number does not remove the misleading story.
WhyIdentifies an audit gap between direct factual probes and downstream decisions, tested with matched controls across eight models and a real bushfire misinformation case.
In briefPsyEvo: A Counseling AI That Personalizes and Improves Itself During Use
SummaryThis paper introduces PsyEvo, a large language model-based counseling framework that adapts to individual clients and improves its response policy at test time. It combines three components: a hierarchical Bayesian skill policy, inter-session listwise preference optimization, and state-conditioned ordinal credit assignment. In simulated-client evaluation on PsychEval, PsyEvo achieves an overall score of 7.684 and outperforms every component variant in three matched runs.
In briefDyRA: Dynamic Residual Approximation for Faster Matrix Multiplication in Neural Networks
SummaryThis paper proposes DyRA, a method that speeds up matrix multiplication in large neural networks by correcting output errors caused by structured weight approximations, using the input at inference time. Unlike prior methods that approximate only the weights, DyRA targets the output activations that determine accuracy. The authors report improved accuracy-efficiency trade-offs across vision, speech, and language models, and a 1.5x end-to-end GPU speedup for DINOv3 with over 3x less accuracy degradation than weight-only baselines.
In briefTopology-driven binding for omni multimodal graph foundation models
SummaryThis paper proposes GraphBind, a method that uses graph topology to bind information from multiple modalities into a unified shared space, targeting graphs with incomplete node attributes. It organizes each node's own semantics and reliable neighborhood semantics through the topology and adapts the shared space to discriminative and generative tasks via lightweight interfaces. According to the paper, GraphBind outperforms 11 representative baselines on both task types, with relative improvements of up to 28.1% over the strongest baseline.
In briefEvidence Found but Not Read: How Extracted Text Bridges the Retrieval-Reading Gap in Document VLMs
SummaryThis paper examines retrieval-augmented document question answering, where a vision-language model may fail to use evidence even after the correct page is retrieved. Using a paired protocol on FoveDoc-Bench, the authors compare answers from page images alone with answers from the same images plus extracted text. Adding CPU-OCR text improves strict accuracy across six VLMs. The gain helps on textual evidence but is neutral or harmful on charts and figures, and it shrinks as retrieval degrades.
WhyIsolates a retrieval-reading gap via a paired protocol, validated across six VLMs with clear boundaries, offering fresh insight mainly for document QA.
In briefTesting how vision-language models answer questions that break conversational rules
SummaryThis paper examines how Vision Language Models perform on visual question answering when questions violate Grice's maxims, such as by adding unnecessary, ambiguous, or false information. The authors show that models including ChatGPT, Claude, Gemini and Llava perform worse under these violations. The study also compares pragmatic reasoning between humans and models, and finds that humans expend less effort resolving model-induced violations, though the models themselves are less accurate in those cases.
Monika Shah, Sudarshan Balaji, Somdeb Sarkhel et al.
In briefBeyond Score Agreement: Analyzing Residual Judging Difficulty in LLM Evaluators with Psychometrics
SummaryThis paper examines the reliability of large language models used as automatic judges, going beyond the usual check of agreement with human scores. Using SummEval summarization evaluations, the authors apply Many-Facet Rasch Models to separate ratings into latent summary quality, rater and dimension severity, and rating-scale thresholds. They define residual hardness to compare which cases are difficult for humans versus 17 open-weight LLM judges. They find that moderate agreement on latent quality does not mean shared residual hardness, and that the mismatch depends on the dimension, with consistency skewing toward LLM-hard cases and coherence toward human-hard cases.
Longwei Cong, Sonja Hahn, Sebastian Gombert et al.
In briefCan See and State, But Not Use: Gap Between Reportable Spatial Facts and Usable States in Multimodal LLMs
SummaryThis paper shows that multimodal large language models can correctly report a spatial fact without using it in later reasoning. The authors introduce SpaceConflict, a benchmark of 23,196 inputs spanning four levels from local fact binding to state judgment under transformation. Models recover initial states from visual evidence but often fail when that state must drive a transformation. The authors propose Operational State Supervision (OSS), which improves paired accuracy most on cross-observation consistency and transformation-based judgments.
In briefQuery-aware routing to improve cross-lingual retrieval in multilingual encoders
SummaryMultilingual encoders often retrieve documents less effectively when queries and documents are in different languages. The authors train a query-only low-rank adapter (SampoTron) against frozen document embeddings and route queries deterministically by language, sending cross-language queries through the adapter and same-language queries through the original encoder. On a sampled financial benchmark across six English, Finnish, and Swedish directions, average nDCG@10 rises from 0.241 to 0.291, a 20.9% relative gain. All six cross-lingual directions improve while same-language performance is preserved.
In briefSceneFactory-3D: Physics-based simulator testing how road friction changes driving safety
SummaryThis paper presents SceneFactory-3D, a GPU-batched multi-agent driving simulator in which vehicles follow acceleration and steering commands through tire-road friction and suspension forces. The authors run matched counterfactual worlds that differ only in road condition, testing three learned-policy families and two classical planners across 21 friction and grade conditions. When friction drops from 1.0 to 0.18, the share of vehicles clearing a work zone safely falls by 6 to 90 percentage points for learned policies and 18 to 19 for classical planners, and near-collisions become more frequent for every learned policy.
In briefConvoDrift: A multi-turn dataset for studying how conversational tone shifts over time
SummaryThis paper introduces ConvoDrift, a dataset for modeling how the stylistic tone of a conversation drifts progressively while the semantic intent stays fixed. It contains 15,727 multi-turn conversation structures, each with six prompt-response pairs annotated for style drift and style direction. The authors also build a pairwise preference dataset with persona-conditioned labels to support research on personalisation and pluralistic alignment. Human and automatic evaluations report high annotator agreement and show that drift changes wording while preserving semantic similarity.
Vihindi Kotalawala, Pamoda Dilranga, Gayani Thoradeniya et al.
In briefPeer effects in signed networks: separating influence from supportive and hostile ties
SummaryThis paper studies how interventions affect people through social networks that contain both positive and negative relationships. The authors define effects for positive ties, negative ties, and their interaction, and propose SiDE, a doubly robust estimator. Semi-synthetic experiments on six real signed networks show accurate effect estimation. An exploratory reanalysis of published school-experiment data gives a positive peer-effect estimate, but all four effect intervals include zero after multiple-comparison adjustment.
In briefRobust Survival Models Against Subgroup Shift and Outlier Contamination
SummaryThis paper proposes a distributionally robust framework for survival analysis that handles latent subpopulation shift and outlier contamination in training data together. An outer minimization reduces the influence of contaminated samples to refine the nominal distribution, while an inner maximization targets the most challenging subpopulation. Experiments on simulated and two survival benchmarks show more stable training and substantially better worst-group performance, with competitive overall results.
In briefTACD: Making text-to-motion models faster via terminal amplification control distillation
SummaryThis paper presents TACD, a distillation method that trains efficient motion generators from text prompts and pretrained teachers without real motion data. It ties the teacher query to the student's step size to bound effective loss weights in clean-motion space, which improves few-step generation. The authors report a 58% FID reduction for an eight-step HY-Motion student on HumanML3D relative to distillation without the bound, and 7.7 to 11.9 times end-to-end speedups on HY-Motion and Kimodo.
In briefDIVINE: Cross-market stock pretraining by reconstructing diverse technical indicators
SummaryThis paper proposes DIVINE, a pretraining framework that learns financial time-series representations by reconstructing technical indicators computed from raw price-volume (OHLCV) history. It is pretrained jointly on six equity market datasets to reconstruct 77 targets derived from 16 standard indicators, and only the encoder is transferred to downstream stock ranking. The authors report that a lightweight 0.05M-parameter encoder achieves the strongest average portfolio performance across all six markets.
In briefUnlearning for Time-Series Forecasting: Removing Deleted Data's Influence from Forecast Models Without Full Retraining
SummaryThis paper studies how to remove the influence of deleted training data from time-series forecasting models without full retraining. It proposes RDTU, which first uses a retained-set neural tangent kernel predictor to get a base forecast, then measures each affected window's structural support using retained reference data. A diffusion model generates a residual correction that forms a pseudo-label guiding a lightweight model update. Experiments show the unlearned models most closely match exact retraining.
In briefNeuroLens: Learning Stable Neural Meaning Embeddings from Long-Term Brain Recordings
SummaryThis paper introduces NeuroLens, a self-supervised model based on the JEPA framework that learns denoised, semantically informative latent representations from chronic neural recordings. An adaptive encoder maps changing neural populations into a shared latent space, and a temporal predictor forecasts future latent states. Tested on intracortical data from mice and humans, it improved decoding of decision-making and semantic task variables and gave more stable decoding over time than state-of-the-art baselines.
In briefTesting whether world models truly distinguish action effects, beyond low prediction error
SummaryThis paper argues that low latent prediction error does not prove a world model distinguishes the consequences of its actions. Using a controlled deformable-physics simulator, the authors trace the same intervention through simulator state, raster observations, target embeddings, and predictor outputs. They find that changed commands alter particle motion yet 41.5% of one-step raster pairs are identical, and the predictor responds far less to actions than the target embeddings change. The authors recommend auditing physical effect, observation visibility, representation geometry, and action dependence separately.
In briefHarness-Aware Distillation for Small Agents: Teaching Small Models What Teachers Know Beyond the Harness
SummaryThe paper proposes Harness-Aware Distillation (HAD), a method for distilling language model agents into smaller models while the surrounding software harness stays fixed. HAD focuses on the abilities the teacher has beyond the harness, using action preferences that contrast teacher actions with and without harness information, and it removes pairs that conflict with harness records. Across several long-horizon agent benchmarks and models, HAD outperforms on-policy distillation baselines, entering fewer unproductive loops and recovering from errors more often.
In briefAdaptive Mutual Distillation: Balanced Multi-Task Post-Training for Large Language Models
SummaryThis paper addresses multi-task post-training of large language models, where tasks have unequal amounts of training data. It proposes Adaptive Mutual Distillation (AMD), which jointly trains two models using different task-balancing strategies and selects distillation weight adjustments per task and transfer direction using short probes and validation scores. Across six benchmarks and three LLM backbones, both AMD models outperform supervised fine-tuning baselines on average, and merging the two models yields a single model that beats multi-task SFT by an average of 2.91 points.
In briefCertifiable jailbreak detection with contraction-constrained state space models
SummaryThis paper studies whether a state space model (SSM) safety classifier head can be formally certified to give the same prediction for all inputs within a bounded embedding perturbation. The authors show that certification via interval bound propagation is possible when the state transition matrix's infinity norm is below 1, and enforcing this contraction raises the certified fraction on toxic comment data from 41% to 59%. On jailbreak detection, the head trained on JailbreakBench transfers zero-shot to AdvBench and HarmBench, matching a logistic regression probe on Mamba embeddings, with its main contribution being formal certification.
In briefLow Permutation Error Isn't Enough: Action Collapse in Multi-Agent Transformer Policies
SummaryThis paper studies how multi-agent transformer policies for cooperative navigation are affected when agent order is permuted, a mismatch between unordered teams and ordered token sequences. The authors show that low permutation error can be misleading, since policies may look robust when all agents pick the same action. They therefore add action-collapse diagnostics such as action diversity and same-action fraction. A weak equivariance penalty improves robustness while keeping action diversity for three-agent teams, while four-agent teams need much smaller regularization weights.
In briefTurnover-Orthogonal Credit Assignment (TOCA): separating team member changes from agent actions in multi-agent learning
SummaryThis paper studies cooperative multi-agent reinforcement learning in open teams, where agents can join, leave, or be replaced during an episode. The authors propose TOCA, a value decomposition that separates action effects, pure turnover effects, and their interaction, so agents are not rewarded or penalized for turnover outside their control. Controlled experiments and a replacement-only Dynamic Spread benchmark show TOCA improves returns over event-aware MAPPO-style critics, and removing the interaction credit term substantially hurts performance.
In briefUnderstanding Enrichment in Reinforcement Learning: How hints and importance-weight correction affect training when rewards are sparse
SummaryWhen rewards are sparse, reinforcement learning with verifiable rewards often uses hints to produce more successful rollouts, which biases policy-gradient updates unless corrected. The paper shows that omitting or truncating importance-weight correction implicitly reweights the reward and decomposes the resulting gradient error into scale, rotation, and variance. It also proposes a sequential Monte Carlo weight correction that turns the exponential compounding of correction variance into additive accumulation. The authors apply this analysis to fine-tuning Qwen3-1.7B on a sparse band of OpenMathReasoning, illustrating how enrichment, corrected or not, helps avoid collapse.
In briefDNAlign: A Dynamic Null-Space Method for Making Language Models Safer
SummaryThis paper proposes DNAlign, a lightweight framework for aligning large language models toward safe behavior. It combines control-theoretic optimization with null-space projection, restricting perturbations to a harm-related subspace to preserve general knowledge. Evaluations across multiple LLM backbones report fewer harmful outputs while maintaining fluency, coherence, and factual utility.
In briefHow Robust Multimodal Claim Verification Is When Text Is Rewritten by LLMs
SummaryThis paper examines whether stylistic changes that LLMs introduce into text affect model decisions on scientific tasks. It focuses on multimodal claim verification, which asks whether a textual claim is grounded in given evidence. The authors test 11 open-weight vision-language models from five families, ranging from 2B to 38B parameters, using natural LLM rewriting and controlled injection of a single LLM-associated word. Most models show no significant accuracy drop, though hedging-oriented rewrites cause consistent and significant probability shifts across nearly all models.
In briefPointWAM: 3D World Action Model for Dexterous Robot Manipulation
SummaryThis paper introduces PointWAM, a 3D world action model that splits the world into a scene and hands, then jointly forecasts both as 3D point trajectories in a shared space-time frame. Given a colored point cloud and a language instruction, it predicts how the scene and hands evolve, then retargets the hand motion to robot actions. The authors report that pre-training on human videos and adding scene-trajectory supervision both improve DexJoCo success rates. They state that PointWAM surpasses prior state-of-the-art results on ten DexJoCo tasks and outperforms strong VLAs on a real robot.
Chunghyun Park, Beomjun Kim, Seungcheol Park et al.
In briefBeyond Data Distribution: System Behavior, a Causality Tax, and a Non-causal Base Model
SummaryThe paper proposes the SBD framework, which treats system behavior S, factors beyond the data space, as an irreducible term in the evidence lower bound. It argues that causality may be neither necessary nor optimal, and introduces a "Causality Tax": ignoring S makes causal modeling a suboptimal approximation with extra structural error. The authors also present Green Shell (GSH), a non-causal variational family, and report via NTK analysis tighter error bounds and a signal-to-noise improvement of over 7 dB.
In briefAll Work And No Play: Understanding and Preventing Strategy Collapse in RLVR for LLMs
SummaryThis paper studies a late-stage collapse seen in GRPO-style algorithms when post-training large language models with Reinforcement Learning with Verifiable Rewards (RLVR). The authors argue it is a harmful loss of effective strategy capacity rather than benign pruning. They present a theoretical framework and a lightweight online warning signal called the Mirrored Entanglement Index. They also propose Mesh Learning, which keeps multiple reasoning strategies active, and report gains on several math, science, and coding benchmarks.
In briefFastOPD: On-Policy Distillation for Making Large Robot Action Models Lightweight
SummaryLarge vision-language-action (VLA) models are powerful but too computationally expensive for real-world robot deployment. This paper proposes FastOPD, which distills a large teacher model into a compact student through efficient on-policy distillation, using a flow map for single-state supervision and a self-consistency objective. On LIBERO, the method keeps 84% of π0.5's performance with only two inference steps, cutting inference latency by 78.1%.
In briefAMBER: Budget-Efficient Multi-View Reranking for Vision-Language Models
SummaryVision-language models can rerank candidate lists for multimodal retrieval, but their high inference cost limits how many candidate views they can evaluate. AMBER is an online framework that uses continuous Elo updates to maintain a global ranking and dynamically allocates VLM calls to ambiguous candidate views and informative queries. Experiments on CIRR, CIRCO, and PhotoBench report that AMBER performs best overall among the compared multi-call methods under comparable VLM-call budgets. The paper also reports effectiveness at lower budgets.
In briefClinical Concept Centers in LLMs: Locating and Validating Clinical Concept Representations in Latent Space
SummaryThis paper asks whether clinical concepts exist as locatable, causally used representations inside open-weight large language models. The authors find dedicated clinical concept centers in all eleven open models tested, which fire only on aligned clinical narratives and causally influence behavior in constrained and open-ended settings. Steering models along these centers yields downstream improvements, and a blinded clinician validation finds that their activation and use predicts clinicians' preferences.
Aishik Nagar, Abhishek Vaidyanathan, Arun-Kumar Kaliya-Perumal et al.
In briefFSPO: Policy-Consistent Risk and Pareto-Feasible Control for Budgeted LLM RL Post-Training
SummaryThis paper introduces FSPO, a feedback-state controller that jointly adjusts several training parameters, such as rollout temperature, group size, clipping, and KL regularization, during reinforcement-learning post-training of LLMs. It combines a risk model that follows the deployed controller's behavior, a calibration method (DCTC) based on cross-fitted trajectories, and a certificate (PRCC) that checks whether multiple resources remain feasible over the remaining horizon. Under a matched GRPO resource budget, FSPO reports higher held-out and out-of-distribution accuracy than the strongest evaluated adaptive baseline, PB2.
In briefMLCommons Jailbreak Benchmark v1.0: A Method for Testing How Well Language Models Resist Single-Turn Jailbreak Attacks
SummaryThis paper presents the MLCommons Jailbreak Benchmark v1.0, an end-to-end methodology for evaluating how robust large language models are against single-turn, text-based jailbreak prompts. It covers system and attack selection, paired baseline and adversarial evaluation, human annotation, evaluator calibration, and scoring. The benchmark tests eight open-weight systems with 264 seed prompts and representative attacks. Across the evaluation, the unsafe-response rate rose from 11.08% at baseline to 18.65% under jailbreak conditions, an average Resilience Gap of 7.57%.
Carsten Maple, Cagatay Yucel, Isaac Holeman et al.
In briefExpanding Training Trajectories for Complex Tasks via Recursive Self-Rewrite
SummaryThis paper proposes Recursive Self-Rewrite (RSR), a framework where one base model finds successful solutions to difficult terminal tasks under diverse harnesses and rewrites them as training trajectories for a general harness. Starting from 2,001 successful source trajectories, RSR produces 11,094 rewritten trajectories for supervised finetuning. Models trained on them outperform both the base model and direct trajectory SFT on several terminal benchmarks.
In briefTerrainForge: Physics-Based Road Geometry Editing for Counterfactual Driving Scenes
SummaryThis paper presents TerrainForge, a framework that generates counterfactual driving scenes by editing road geometry such as crests, sags, speed humps, and friction on reconstructed multi-vehicle episodes. A unified road model links these edits to four-wheel vehicle dynamics, camera viewpoint, and inter-vehicle clearance, with checks against CarSim and Waymo scenes. The authors also train a screening surrogate that predicts how a road edit changes the ego vehicle's terminal gap, allowing candidate edits to be filtered before full simulation.
In briefMetaRubric: A Reward-Learning Method for Rubric-Based Reinforcement Learning
SummaryThis paper addresses a problem in rubric-based reinforcement learning, where judges may award high scores even when a response lacks the required information, a failure called Vacuous Credit. The proposed MetaRubric method alternates evidence-aware policy optimization with rubric adaptation guided by current responses, and uses counterfactual prompts that change one task-relevant fact. Across multiple backbones, it improves PubMedQA accuracy by 6.00 to 20.40 percentage points over static-judge GRPO, with further gains on HealthBench-Hard and two multimodal medical benchmarks.
In briefAdaptive spectral-Koopman modeling for handling data shifts over time
SummaryThis paper introduces AdaSpecK, a framework for temporal domain generalization, where streaming data distributions change over time. It uses spectral-regularized Koopman dynamics in latent space to filter noise and learn linearized system dynamics, and a target-conditioned attention module with a learned router to select useful historical context. The authors report state-of-the-art results on eight classification and regression benchmarks.
In briefText-centric training to improve audio-visual reasoning in omni-modal AI models
SummaryThis paper studies how to improve joint audio-visual reasoning in Omni Large Language Models while reducing training cost. The authors find that models can fail multi-hop reasoning even when they answer all related single-hop questions correctly. Text-only reasoning training improves performance but degrades perception, so they propose a text-centric method: text training first, then reduced-data audio-visual reinforcement learning to restore perception. The refinement stage keeps 93.5% of the text-only reasoning gain while using about 90% fewer input tokens than full-data audio-visual RL.
In briefGated Slot Attention-2: Correcting Key and Value Memory in Efficient Linear Attention
SummaryThis paper addresses how linear attention models manage their fixed-size recurrent memory. It introduces Gated Slot Attention-2 (GSA2), which combines a key-side correction rule with a value-side delta-rule correction through shared latent slots. The authors also provide a parallel chunkwise training algorithm. Experiments reportedly show consistent gains over strong linear-attention baselines while keeping linear-time sequence modeling and constant-memory decoding.
In briefiS-KV: Online Low-Rank KV Cache Compression via Block-Incremental SVD
SummaryLong chain-of-thought reasoning makes the key-value cache grow with every decoded token, increasing memory use. This paper proposes iS-KV, which keeps recent tokens exact and incrementally folds older states into bounded-rank representations, synchronizing historical coordinates as the basis evolves. On DeepSeek-R1-Distill-Llama-8B it reaches 82.6% accuracy at 4.06-fold compression, versus 83.6% for the original model, and on Qwen3-8B it reaches 89.2% at 5.64-fold compression. The authors report that under matched memory budgets it outperforms token-eviction baselines.
In briefROUTEAUDIT: A framework for fairly identifying the effects of routing choices across multiple verifiers under budget limits
SummaryThis paper studies how adaptive systems that route tasks among several verifiers should be compared, arguing that endpoint quality-cost gaps can hide changes in catalog, accounting, or scoring. It proposes ROUTEAUDIT, which uses a contract lattice, a policy-independent response tape, and request-level bounds to produce an attribution certificate for each comparison. On 1,319 held-out task requests, the learned and RLVR policies scored 0.9522 and 0.9553 against 0.9484 for a matched static policy.
In briefVIGOR: Zero-shot visual generalization for model-based reinforcement learning
SummaryThis paper proposes VIGOR, a framework that helps model-based reinforcement learning agents perform well when visual conditions change, such as backgrounds, lighting, or camera position. It combines asymmetric weak-to-strong augmentation, dynamics-level latent consistency, and encoder stabilization. The authors report that VIGOR outperforms prior model-free and model-based baselines on DeepMind Control Suite and Robosuite. Ablations indicate that the robustness comes from latent-space consistency rather than the specific augmentation used.
In briefBitNest: Bit-Nested Speculative Decoding to Speed Up LLM Inference With Less Memory
SummaryThis paper proposes BitNest, a framework that embeds a low-precision draft model inside the higher-precision target model so both share one physical weight representation, avoiding extra memory overhead. The same progressive-precision design is extended to the KV cache to support long-context inference. Across multiple 7B–8B edge-friendly LLMs, the authors report an average speculative acceptance rate of 95.2% and a 1.48–1.61x end-to-end speedup over FP16 autoregressive decoding.
In briefFUSEye: Lightweight Fisheye Object Detection Using Overlapping Views and Zero-Initialized Adapters
SummaryFisheye cameras offer mobile robots a low-cost wide view, but COCO-pretrained detectors perform poorly on their distorted images. FUSEye adapts a frozen-backbone YOLO26-x detector by adding about 227k parameters across input, feature, and decision levels, using overlapping grid views, zero-initialized residual adapters, and cross-view agreement fusion. On the WoodScape benchmark, it raises mAP50 from 0.148 to 0.266, retaining 84.3% of full fine-tuning accuracy. Using only 25% of labeled images, it still achieves 97.6% of its full-label performance.
In briefMuon Learns Facts Better: How Spectral Orthogonalization Shapes Feature Learning
SummaryThe paper studies how the spectral orthogonalization used by the Muon optimizer affects feature learning, using a tractable factual-recall model with a linear transformer. It finds that spectral gradient flow reduces the learning-time ratio between subject- and relation-dependent information from order √(S/R) to order 1, and decays errors faster than standard gradient flow. It also shows that spectral and standard gradient flow are equivariant under orthogonal transformations of token embeddings, while Sign gradient flow, the continuous limit of Adam, is not.
WhyUsing a tractable factual-recall model, it proves spectral orthogonalization shrinks the learning-time ratio from √(S/R) to Θ(1), a clear theoretical advance.
In briefCross-subject emotion tracking from EEG and fNIRS brain signals without labels
SummaryThis paper studies predicting moment-by-moment valence and arousal for people whose emotion labels the model never sees, using EEG and fNIRS recordings. The approach splits the emotional trajectory into a structure shared by viewers of the same video and an individual scaling estimated from a label-free EEG alpha-band synchrony measure. The authors report lower error on held-out subjects than the EEGNet and ASAC-Net baselines for the same split. They also report negative results, finding no other signal that improved on the single synchrony marker.
In briefEfficient memory crystallization for graph learning under shifting data distributions
SummaryThis paper addresses graph learning models that must adapt as graph distributions drift over time. It proposes EMC, a training-free test-time framework that compresses each incoming graph domain into a compact memory via a closed-form distribution-matching objective, and models dependencies between domains with state-evolving memories. The authors report a tighter generalization error bound than direct adaptation. Experiments show better performance than state-of-the-art baselines, with 87.4% lower average runtime and 92.4% lower GPU memory use than a recent competitor.
In briefPAPER2LLM++: Letting LLMs Continually Self-Evolve from Research Papers
SummaryThis paper introduces PAPER2LLM++, a framework that lets large language models keep learning from newly published research papers. For each paper, it extracts evidence-grounded findings, checks whether the reported limitation still appears in the current model, and converts findings into candidate learning signals when needed. An update is kept only if it improves the targeted behavior without substantially forgetting earlier gains or degrading general capabilities. Experiments on a sequential stream of research-discovered failures show models can progressively incorporate new findings while retaining prior improvements.
In briefLaw And Order: Automatically Converting Tax Law into Verifiable Code
SummaryThis paper studies how to automatically translate tax forms and filing instructions into executable symbolic programs. The authors propose Law&Order, a neuro-symbolic framework that combines large language model synthesis with cell-level verification and iterative localized error repair. On 51 held-out TaxCalcBench returns that were never used during generation or repair, it reaches 100% cell-level and form-level accuracy, while the most advanced LLM alone reaches 66%.
Sophia Simeng Han, Yoshiki Takashima, Anjiang Wei et al.
In briefBeyond Token Accuracy: Prioritizing Key Content for Visual Reconstruction
SummaryThis paper proposes AdapToC, an adaptive framework for visual-token communication aimed at image reconstruction quality. The transmitter dynamically decides how many tokens to send and assigns protection levels based on image content, channel state, and budget. The receiver uses an adaptive MaskGIT decoder that accounts for channel reliability to correct and recover missing tokens. Under matched communication costs, it achieves a peak mean PSNR gain of 4.20 dB over the strongest static baseline.
In briefHold-Out Scoring for Efficient Gaussian DAG Learning: a new fast method for learning causal graph structure from data
SummaryThis paper introduces HOST, an algorithm for learning the structure of Gaussian directed acyclic graphs (DAGs) without expensive subset search or a given indegree bound. It orders nodes using hold-out scoring and convex regression, and then recovers parents by removing indirect effects from total effects. Under suitable conditions, the authors report exact recovery of a p-node DAG with maximum indegree d using a sample size on the order of d log p, in polynomial time. Experiments indicate competitive graph recovery and favorable runtime scaling.
WhyReplaces subset search with hold-out scoring and convex regression, reaching near-optimal sample complexity in polynomial time; experiments are modest in scale.
In briefWarm-starting rubric-based RL with on-policy distillation before reinforcement learning
SummaryThis paper proposes a two-stage training framework for language models on open-ended tasks scored by rubrics. In the first stage, a student model without access to the rubric matches a rubric-aware teacher's next-token distributions on its own generated prefixes, providing dense token-level supervision. In the second stage, reinforcement learning optimizes the rubric reward directly. The authors evaluate on HealthBench, ResearchQA, and RubricHub Science with open-weight models, reporting the highest scores among the compared methods, and note that an SFT plus RL baseline increasingly rewards claims of rubric compliance without the required content, a sign of reward hacking that is limited for their approach.
In briefControlling Polar Exposure to Delay Memorization in Diffusion Models
SummaryThe paper studies the window before diffusion models begin copying training images while still producing good-quality samples. It proposes Quality-Gated De-whitening (QGD), which keeps a fast polar-update phase and gradually restores fixed-gain momentum once quality is confirmed, to delay copying. On 2,000-image CIFAR-10 subsets, QGD reports an 8.32x wider useful interval and 75.9% less copying than the baseline, with FID 75.56 versus 79.37 for SGD.
In briefTRAC: A method that speeds up autoregressive video generation while preserving quality
SummaryThis paper introduces TRAC, a framework that makes autoregressive video generation faster and more robust. It combines three components: Robust Cumulative Scheduling for cache reuse, Autoregressive Trajectory-aware Guidance Scheduling, and Spectral Structure Correction for restoring low-frequency structure. On VBench with SkyReels-V2, the authors report a 6.03x speedup over the vanilla model with only about 0.6% score degradation.
In briefBenchmark for Automatically Evaluating Mental Health Stigma Online
SummaryThis paper introduces a theory-based benchmark for automatically evaluating mental health stigma in online news and social media text. The texts mention six mental health conditions and are annotated with a fine-grained taxonomy covering stigma mode, domain, and components. Results show that models for sentiment, toxicity, and hate speech do not capture mental health stigma well, and that large language models often overpredict stigma unless given explicit operational rules.
In briefLatticeSMC: Allocating Inference Compute in Chunked Sequence Generators
SummaryThe paper proposes LatticeSMC, a budget-matched inference-time steering sampler for models that generate long sequences such as music, motion, and video chunk by chunk. It resamples particles on a two-dimensional grid of chunk index and denoising step, using potentials derived from a Feynman-Kac model. Under matched compute on music-to-dance and 40-second text-to-music tasks, it improves beat alignment and prompt adherence at 32 particles while preserving held-out quality, and human raters prefer it in 60-77 percent of pairwise comparisons.
In briefImproving recall of individual facts in unstructured knowledge editing using focused views
SummaryLarge language models may not reflect facts that changed after pretraining, and knowledge editing offers a targeted way to update them. This paper finds that existing unstructured knowledge editing methods suffer from context reliance, where edited models can reproduce a passage but fail to recall its individual facts without the original context. The authors propose FOVEATED, which randomly shifts rotary position embeddings of preceding context keys during editing to build focused views of each sentence, then removes the shift at inference. The authors report consistent improvements across five editors, two LLM backbones, and three benchmarks.
In briefNear-optimal best-arm identification using only 1-bit feedback
SummaryThis paper studies finding the best arm with fixed confidence when the learner receives only one bit per round, indicating whether the sampled reward falls in a chosen query set. The authors build a time-uniform 1-bit mean-estimation method using randomized threshold queries and a clipped tail-integral identity, and embed it in candidate-challenger algorithms. A phased adaptive-clipping algorithm achieves gap-adaptive sample complexity, and a K-arm lower bound shows the logarithmic penalty from 1-bit finite-variance feedback is intrinsic, matching the algorithm's leading dependence up to lower-order log log factors.
In briefAptMQL-Bench: A Text-to-MQL Benchmark Built from Text-to-SQL Data
SummaryThis paper introduces AptMQL-Bench, a benchmark for converting natural language into MongoDB Query Language (MQL). It is built from the BIRD text-to-SQL dataset by designing document schemas around expected access patterns and migrating data from SQLite without loss. The benchmark includes 21 databases, 3,186 natural-language requests, and their MQL queries. The strongest model, Claude Opus 4.5, reaches 57.38% accuracy without external evidence and 70.34% with it, indicating the task remains challenging.
In briefWhen Agents Fail to Learn from Their Own History
SummaryThis paper studies how language agents use past attempts and feedback within the same task. The authors find that interaction history generally helps, but shuffling past actions barely reduces success, suggesting agents do not reliably link actions to their outcomes. Explicitly labeling each observation as the result of the preceding action improves task success and reduces repeated actions. The authors further introduce a learned calibrator that reassesses past actions and selectively records experience, improving success beyond labeling alone.
In briefNo-Free-Graph: Deciding When Multimodal Data Should Be Turned into Graphs
SummaryThis paper finds that adding graph structures to multimodal data does not always help, sometimes giving large gains and sometimes little or negative benefit. The authors propose MAG-SCOUT, a framework that assesses, before building the full graph, whether graph construction is worthwhile, and makes a build-or-skip decision. Across six multimodal datasets and three downstream tasks, it reportedly saves 33.1% of graph construction work while retaining 96.7% of held-out positive-gain cases.
In briefExact memory-time optimization for caching language model prefixes
SummaryThis paper studies how long to keep cached language-model prefix states, balancing recomputation against storage time. The authors propose Prefix-Certificate Retention (PCR), which models usable-prefix rewards as a maximum-weight closure solvable by one minimum cut. Experiments replay 39,632 public Mooncake requests and measure usable prefix blocks and storage time, not GPU latency.
In briefLocal conformal safety monitoring for autonomous driving using vision-language models
SummaryThe paper proposes SLLCP, a post-hoc calibration layer that wraps frozen vision-language models to turn their unreliable collision predictions into probabilistically calibrated safety prediction sets. It applies a localized procedure that upweights relevant past experience based on the observed driving scene when computing uncertainty thresholds. Evaluated on 15k CARLA trajectories from unseen scenarios, SLLCP flagged 89.6% (Qwen) and 88.4% (Cosmos) of collision-causing trajectories, compared with 4.6% and 39.1% for the base models.
In briefA Controlled Test of Whether Personal AI Memory Helps Predict Ratings
SummaryThis paper examines whether a personal AI, in predicting structured ratings, uses past item-rating associations or mainly the user's own rating tendencies. The authors shuffle each user's historical ratings while keeping the rating distribution, item coverage, and metadata unchanged, and evaluate 400 held-out user profiles with 6,160 target ratings on Coat and MovieLens. They conclude that extraction, use of associations, output reliability, and reader choice need separate evaluation.
In briefA two-stage method for near-real-time forest disturbance alerts using Sentinel-1 radar
SummaryThis paper presents a two-stage statistics-encoder cascade for near-real-time forest anomaly detection from Sentinel-1 SAR time series. It addresses cloud cover that limits optical monitoring and seasonal radar changes that can be mistaken for forest loss. A robust z-score test on VH backscatter is confirmed by a convolutional autoencoder SSIM gate, and alerts are issued only when both stages agree, with confidence and risk tiers.
In briefDynamic LLM routers are often misguided: an analysis of six commercial routers
SummaryThis paper evaluates six commercial dynamic LLM routers, which send each query to a model chosen to cut inference costs, across 14 settings and eight task categories. At matched cost, none outperformed a router that randomly chose between two well-selected models, and some trailed by more than 10 percentage points. The authors identify four recurring patterns (difficulty blindness, length reversal, semantic matching, and roster suboptimality) and propose a new evaluation method and a simple two-model router, whose gain over random routing is limited.
In briefExpanding Attention: a Taylor-series method for linear-time Transformers with a fixed number of states
SummaryThe paper expands attention's exponential kernel into a truncated Taylor series and compresses the causal history into a fixed number of recurrent states that does not depend on context length. The authors report a 20.8 times speedup over softmax at N = 8192 and 76.0 times faster per-token decoding at N = 65536. They find that second order (n = 3) gives the best trade-off between error and state size, and that query and key normalization is required.
In briefFiberGeoText: A vision-language model that groups short brain fibers across people
SummaryThe superficial white matter is a brain region with many short connecting fibers whose layout is not yet well understood. This paper introduces FiberGeoText (FGT), a vision-language model that groups short fiber tracts reconstructed from ultra-high-resolution diffusion MRI into population-level clusters. It jointly encodes each fiber's 3D trajectory, cortical anatomical context, and shape, with cortical endpoint descriptions encoded by a pretrained large language model. On 0.76 mm diffusion MRI data, FGT outperformed state-of-the-art methods on parcel coherence, shape and cluster-size consistency, and cross-subject correspondence, and it generalized to unseen subjects.
Yuqian Chen, R. Jarrett Rushmore, Guikun Chen et al.
In briefJumping up and down: denoiser diffusion models for discrete ordinal data, which are whole-number data with a meaningful order.
SummaryThe paper introduces JUD (Jumping Up and Down), a family of denoiser-based diffusion models for discrete ordinal data. It is described as the first ordinal-data diffusion family built around training denoisers, and it allows bi-directional (up and down) perturbations of the data. The authors report that the simple training objective and flexible perturbations yield competitive results across different data modalities.
In briefCorrecting guided diffusion sampling using spectral alignment
SummaryThe paper proposes Spectral Correction Guidance, which uses the spectrum of intermediate diffusion states, compared against an analytic reference spectrum of the forward process, to correct deviations during sampling. The method is training-free and can be applied across diffusion backbones and conditional generation tasks without modifying the model. The authors report gains over baseline guidance methods in text-to-image generation and improved quality over classifier-free guidance on ImageNet, with improvements persisting across guidance scales and fewer denoising steps.
In briefQuery encoders may be hard to learn even when vector retrieval geometry is favorable
SummaryThe paper studies whether learned single-vector query encoders can reach the maximum recall a frozen document index can support. On several real-world retrieval benchmarks, their retrieval quality often falls far below that ceiling. The authors construct a retrieval task where a perfect-recall query encoder exists as a small one-hidden-layer ReLU network, yet any statistical-query learner needs exponentially many queries to beat the random baseline k/n. This suggests query encoder learnability may be a barrier in embedding-based retrieval.
In briefEpiWorld: Grounding LLM Policy Agents in Epidemiological World Models
SummaryThis paper presents EpiWorld, a closed-loop framework that pairs a large language model policy agent with a learned, action-conditioned epidemiological world model and a skill library of public-health protocols, surveillance tools, and lessons from after-action analysis. The world model runs counterfactual simulations of candidate interventions to guide policy selection while protocol constraints remain fixed. On retrospective COVID-19 and Influenza data, the framework reduced cumulative hospitalisation by up to 59% across datasets and by about 16% on average across six LLM backbones, outperforming reinforcement-learning and optimal-control baselines.
In briefMonitoring Reasoning Traces in Looped Language Models
SummaryThis paper provides the first systematic evaluation of how looped language models, which repeatedly apply shared transformer layers, affect chain-of-thought (CoT) monitorability, meaning how well their reasoning can reveal undesirable behavior. The authors compare different loop depths within one model family and compare looped models with non-looped models matched by size, layer count, or effective depth, across eight MonitorBench tasks. Under stress tests, some Logic, Science, and Engineering Cue Answer tasks show reduced monitorability at deeper loop depths, but looped models were not systematically less monitorable than matched non-looped models.
In briefProspective Hindsight: Self-Calibrating Reinforcement Learning from Prediction-Reality Gaps
SummaryThis paper proposes Prospective Hindsight (PH), a training principle that adds to retrospective reinforcement learning methods a signal based on the gap between an agent's prediction before feedback and its evaluation after feedback. Samples with large surprise receive amplified gradient weight, and the predictor shares parameters with the policy so the two co-evolve. The authors report that PH improves both task performance and calibration on single-turn verifiable tasks and a multi-turn personal-agent task.
In briefBeyond Correctness: Handling Unspecified Details in Agentic Text-to-SQL
SummaryAgentic text-to-SQL systems can ask users clarifying questions before generating SQL. The paper argues that a correct execution result does not prove the agent resolved ambiguity, since it may silently make unverified assumptions. The authors propose PlanPool, which treats the clarification plan as a mutable question pool where each planned question must be asked or dropped before submission. Across three benchmarks, PlanPool improves ambiguity coverage and reduces silent failures while keeping execution accuracy competitive.
Wen-Zhi Li, Yue Gong, Konstantinos Kanellis et al.
In briefTPBench: A Benchmark for Dialogue Compression That Loses Turning Points
SummaryThis paper introduces TPBench, a benchmark that tests whether dialogue compressors keep the user turns that revise earlier information, such as a changed price, choice, or constraint. It uses three probes: the user's initial goal (P1), the current value of a revised slot (P2), and a joint probe asking for both (P3). Current-value answers come from existing human dialogue-state annotations in MultiWOZ and SGD, so no new crowdsourcing is needed. On the joint probe at a 0.30 retained fraction, every tested compressed method stays below full context.
In briefMinimizing Bellman error with linear programming normalization
SummaryThis paper proposes a new functional approximation method to reduce Bellman error in high-dimensional dynamic programming and reinforcement learning. It uses network capacity control in revenue management as a motivating example. The authors show that deep neural networks can be combined with linear programming approximation algorithms to derive approximate solutions. Simulation results indicate the proposed algorithms perform competitively against a benchmark.
In briefSymRegFlow: Symmetry-Regularized Flow Matching for Video World Models
SummaryThis paper introduces SymRegFlow, a flow-matching framework that generates multi-view-consistent videos across continuously varying camera viewpoints without ground-truth novel-view RGB supervision. It geometrically warps source views into noisy anchors and combines masked dual-anchor supervision with cross-anchor consistency regularization. On nuScenes, SymRegFlow achieves the lowest FVD and FVMD among the evaluated baselines, reducing FVD by over 31% relative to the best baseline.
In briefGenerating brain structure and function connectivity together using hypergraph-based flow matching for multi-way region interactions.
SummaryThis paper proposes MHG-FM, a framework that jointly generates structural connectivity (SC) and functional connectivity (FC) of the brain and translates between them. It builds modality-specific hypergraphs to capture higher-order interactions between brain regions, encodes them with hypergraph neural networks, and fuses them via dual cross-attention. Conditional flow matching in a variational autoencoder's latent space performs the synthesis and translation. On the HCP-YA dataset, the authors report better reconstruction, topology preservation, distributional similarity and SC-FC coupling than several baselines, with roughly 8x faster sampling than a matched diffusion model.
In briefImproving VLM visual representations with kernel canonical correlation analysis
SummaryVision-language models like CLIP are strong at semantics but weaker at fine-grained visual perception. The paper argues that reducing the alignment loss to DINOv2 does not necessarily hurt fine-grained vision, suggesting kernel-matrix alignment may be insufficient. It proposes aligning feature subspaces with Kernel Canonical Correlation Analysis (KCCA), and extends this to a three-view method, 3vKCCA, that also includes text encoder projections. With CLIP ViT-L/14 on ImageNet-1K, 3vKCCA raises MMVP-VLM accuracy from 17.8 to 25.9 while maintaining zero-shot image-text retrieval performance.
In briefEgo2World: Turning First-Person Cooking Videos into Executable Planning Environments for Belief-State Planning
SummaryThe paper introduces Ego2World, a benchmark that converts annotated first-person cooking activities into executable planning environments under partial observation, so an agent's proposed actions can be run and their outcomes checked. Evaluating six planners on 105 tasks, the authors find that accepted operations often leave task goals unmet. In a paired Qwen-Plus study, persistent belief raised action validity by 4.15 percentage points and cut visual-query attempts by 90.27%, with higher token use and no detected completion gain.
In briefWakeKV: a reversible KV cache residency policy for attention heads that shift behavior during generation
SummaryMost KV-cache compression methods classify attention heads once and keep that classification fixed during generation. The authors measured across three models (1.5B-8B) and three regimes, finding that most heads change their reading behavior at least once during generation. They propose WakeKV, which moves cooling heads to a recoverable CPU reservoir instead of freezing or permanently evicting them, and report lower miss rates than frozen classification and destructive eviction at matched memory or budget. A FlexiCache/vLLM implementation on Mistral-7B shows higher throughput while retaining LongBench quality.
In briefPerformance gap in wearable activity recognition for older adults
SummaryThis paper examines how well wrist-worn accelerometer models that recognize human activity perform for older adults. Using MyMove, an annotated free-living dataset with a mean age of 71, the authors evaluate several deep-learning architectures and training methods. Gains on younger-adult benchmarks do not carry over equally, leaving a persistent and often widening performance gap. Frozen self-supervised features pretrained on the age-diverse UK Biobank substantially improve older-adult performance and narrow the gap, though disparities remain.
Hossein Khayami, Sungjin Hwang, Eshed Ohn-Bar et al.
In briefSelf-Supervised Scaling of Terminal Environments for Scientific Domains
SummaryThis paper introduces Software-in-the-Loop Reconstruction (SWR), a framework that derives reference outputs and verification targets from existing executable scientific software workflows. An agent rebuilds each program from public input-output examples without access to the source, and is checked against hidden configurations. Applied to 500 workflows across six domains, the method yields 3,000 training examples, and fine-tuning Qwen3.8-27B raises mean Terminal-Bench 2 performance from 47.94% to 53.56%.
In briefMuonIO: A norm-aware optimizer extending Muon to embedding and output layers
SummaryThe paper proposes MuonIO, a single Muon-style update rule applied to both the embedding table and the language model head, which standard Muon implementations handle with AdamW instead. The authors choose different operator norms for the two layers based on their input and output geometries, and show these reduce to a shared vocabulary-oriented normalization. In 1B LLaMA pretraining on C4, MuonIO reportedly cuts I/O optimizer state memory by 50% and I/O update FLOPs by about 46% compared with Muon, while also improving validation perplexity.
In briefLabel-efficient time series classification with a dual-stream OSSE-LSTM and counterfactual attribution
SummaryThis paper studies how many labeled examples per class are needed to build a usable time series classifier when labels are costly. The authors propose Dual-Stream OSSE-LSTM, combining an Omni-Scale CNN with squeeze-and-excitation and a bidirectional LSTM, and introduce Counterfactual Integrated Gradients (C-IG) for explanations. On 19 univariate UCR datasets, it achieves the highest average accuracy and per-dataset win count at every support size.
In briefLearning to Revise Reasoning via Segment-wise On-Policy Distillation
SummaryThis paper addresses limitations of on-policy distillation (OPD) for large language model reasoning, where token-level supervision does not offer a coherent alternative reasoning step and can reinforce poor prefixes. The authors propose Seg-OPD, which selects student reasoning segments using an uncertainty metric, obtains teacher redrafts, and trains the student to prefer them while keeping token-wise supervision. Experiments on mathematical reasoning and competitive programming report higher revision success rates and an average relative accuracy improvement of 5.22% over baselines.
In briefSilent Dissent: LLM agents that yield to majorities still internally hold their original answer
SummaryThe paper studies why LLM agents in multi-agent debate often adopt a unanimous majority's answer. Using two-hop factual questions whose intermediate entity is never stated, scripted peers assert a wrong answer, and the authors read the agent's internal representations with the Jacobian lens and logit lens. Agents that changed their answer still represented their original intermediate entity internally, and hiding the earlier answer altered conformity rates.|
WhyIntroduces residual-stream bridge reading to separate stated conformity from internal belief, with pre-registered multi-model tests and controls.
In briefCapacity and routing trade-offs in Mixture-of-Experts Particle Transformers for jet classification
SummaryThe paper compares dense and Mixture-of-Experts Particle Transformers on the 188-class JetClass-II dataset, varying expert count, routing capacity, top-K, and auxiliary loss. When token dropping is avoided, top-1 MoE models outperform the dense baseline at nearly unchanged nominal compute, while adding more stored experts gives little extra accuracy. Activating multiple experts per token yields further gains at higher computational cost. Routing analyses show expert assignments sometimes correlate more strongly with particle identity and kinematics, but this structure does not increase monotonically with performance.
In briefAdaptive Iterative Repair for On-Policy Distillation: Learning from Evolving Errors
SummaryThis paper introduces AIR-OPD, a framework for on-policy distillation that repairs a student model's errors iteratively. After a failed response, a guidance generator produces repair guidance for the current error, and the student retries using that guidance, with new guidance generated if the retry still fails. Trained on DAPO-Math-17K and evaluated on AIME24, AIME25, and HMMT25, it reports gains of up to 3.6 points over the strongest baseline and preserved base-model performance on MMLU-Pro and GPQA.
In briefSTRIDE: A Method for Tracking Glioblastoma Changes Across Brain MRI Scans Over Time
SummaryThis paper proposes STRIDE, a framework that takes two post-treatment brain MRI scans and the interval between them to classify glioblastoma as stable disease, pseudoprogression, or true progression. It combines lesion-guided spatial representation with an interval-conditioned estimate of change between visits, fused with the observed follow-up scan. On the Burdenko three-class task it reports a macro ROC-AUC of 0.816 and a macro F1-score of 0.796.
In briefGeoScaffold: Training Compact 3D-Aware Latents for Efficient Vision-Language Navigation
SummaryThis paper proposes GeoScaffold, a training-time framework that improves streaming Video-LLM navigation policies for vision-and-language navigation (VLN). It learns a depth tokenizer and trains learnable geometry query tokens to reconstruct depth, connectivity, and traversability. The auxiliary components are discarded after training, so inference needs no depth sensors or perception tools. The authors report consistent gains over leading vision-only navigators on continuous VLN benchmarks.
