In briefContinual learning for robot grasping from experience and human demonstrations
SummaryThis paper presents a continual-learning framework for synthesizing 6-DoF grasps with a parallel-jaw gripper from a single view in cluttered scenes. Instead of fine-tuning a large parametric model, it adapts through memory in a learned embedding space, where grasp outcomes update future scores and optional user demonstrations are transferred to new scenes as candidate grasps. The method was evaluated in simulation and in over 1500 real-world grasp trials.
WhyContinual grasp learning via embedding-space memory avoids finetuning, validated over 1500 real-world trials, offering useful advances within the subfield.
Giulio Schiavi, Andrei Cramariuc, Michael Pantic et al.
In briefQuestionnaire-guided splitting of household smart meter data into appliance-level energy use
SummaryThis paper presents a label-free system that breaks 30-minute household electricity data from Irish smart meters into nine appliance categories, combining high-power event detection with questionnaire-guided estimation. Evaluated on four datasets against two independently developed disaggregation systems, the hybrid method achieved the lowest whole-decomposition error across all buildings. The authors state it also gives useful advice on household consumption patterns and on reducing or shifting use of some appliances to lower costs.
Achal Nanjundamurthy, Rupam Misra, Suzanne Little et al.
In briefITC-MoE: Importance-Guided Token-Aware Compression for MoE Diffusion Language Models
SummaryThis paper addresses the high computation and storage costs of Mixture-of-Experts diffusion language models, whose many expert parameters make them expensive to run. It proposes ITC-MoE, which combines importance-guided adaptive Tucker compression that allocates ranks under a fixed parameter budget, and token-aware compensation that adds low-rank correction for sensitive hot tokens and limits expert choices for cold tokens. The authors report that on SDAR-30B-A3B-Chat-b32 under a 30% compression budget, the method keeps 96.33% accuracy on MultiArith and reaches up to a 7.22x end-to-end speedup.
In briefODDR: One-step shadow removal via reward-guided diffusion, without real paired data
SummaryThis paper proposes ODDR, a shadow removal framework that does not require real-world paired training data. It first trains a one-step diffusion baseline, ODD, on synthetic shadow data, then fine-tunes it with ShadowReward, a reward model trained without human annotation to rank images by perceptual quality. According to the authors, ODDR narrows the gap to fully supervised methods while remaining efficient as a single-step model.
In briefTraining-free 4D reconstruction of dynamic scenes using depth foundation models guided by vision-language models
SummaryDyna3 is a training-free framework that extends the Depth Anything 3 depth model to reconstruct dynamic 4D scenes without fine-tuning. It uses cross-view features with best-match search to separate static surfaces from moving objects, and vision-language models to generate prompts for SAM 3 instance segmentation. The scene is decomposed into an aligned static background and per-frame dynamic point clouds. It outperforms the correspondence-trained VGGT4D on dynamic object segmentation and runs faster with lower memory use.
In briefModel Validation in Machine Learning: A Scenario-Based Guide to Nested Group Cross-Validation
SummaryThis tutorial reviews model validation methods, from hold-out splits and repeated subsampling to k-fold, leave-one-out, group-aware, and nested group cross-validation, with examples from biomedical data. Eight controlled scenarios compare flawed designs against leakage-safe ones, covering feature selection, normalization leakage, dependent records, and multicenter mixing. The paper also provides a data-size matrix, a decision tree, reporting checklists, and reproducible MATLAB and scikit-learn templates. It concludes that no validation method is universally best, that the independent unit must match the deployment target, and that all data-dependent operations must exclude observations used for performance estimation.
In briefShelfChange3D: Object-Level 3D Change Detection for Retail Shelf Monitoring
SummaryThis paper frames retail shelf monitoring as object-level 3D change detection, finding changed products between two RGB-D observations and localizing them with 3D boxes. The authors introduce ShelfChange3D, a dataset of 145K synthetic and 5K real-world paired RGB-D observations with object-level 3D change annotations. They also propose ChangeBox, an end-to-end framework with a geometry-based refinement stage using depth and gravity priors. Experiments show it outperforms existing baselines and transfers effectively from synthetic to real-world data.
In briefA review of trustworthy DataOps and MLOps for intelligent transport and logistics
SummaryThis paper reviews trustworthy Data Operations (DataOps) and Machine Learning Operations (MLOps) in intelligent transportation and logistics. It identifies gaps in existing literature and discusses the necessity, key components, tools, practical insights and case studies of these practices. It also examines methods for strengthening trust in AI systems and concludes with remaining challenges and future directions.
Antonio Emanuele Cin\`a, Giovanni Scodeller, Cecilia Caterina Pasquale et al.
In briefPickMoment: Learning Continuous-Time Frame Recovery from Motion Blur in a Single Image
SummaryThis paper presents PickMoment, a model that learns the average motion blur over any sub-interval of a camera exposure, instead of predicting only a single sharp frame. One deterministic network handles single-image deblurring, blur-to-video generation, and recovering the frame at any chosen moment. The authors train it with three losses derived from the blur integral, and report state-of-the-art results among generative deblurring methods on GoPro and HIDE, competitive results on RealBlur, and the highest per-frame fidelity on GoPro-7 blur-to-video, all in a single forward pass.
In briefSCOPE-AD: Cost-aware sequential diagnosis planning for Alzheimer's disease using energy-based models
SummaryThis paper presents SCOPE-AD, a framework for diagnosing cognitively normal, mild cognitive impairment, and Alzheimer's disease cases by sequentially choosing which tests to acquire under cost and burden constraints. An energy-based teacher's action distributions are distilled into a Qwen policy that decides between further acquisition and diagnosis. On ADNI, it reaches 77.70% Macro-F1 at an average acquisition cost of $50.46, and the authors argue selective acquisition is more cost-effective than full-modality evaluation.
Ziwen Yu, Ivan Koychev, Elizabeth Coulthard et al.
In briefKeeping Correct Tokens: A Self-Correction Gap in Diffusion Language Models
SummaryThis paper finds that uniform-state diffusion language models can revise any token but fail to retain correct ones, revising hundreds of positions per step and reducing sample diversity. The authors propose CTR-Reg, an auxiliary loss that trains the model to keep tokens left unperturbed by the forward process, with no change to the sampler. CTR-Reg raises clean-token accuracy by 26.5 percentage points on average across six benchmarks and more than halves generative perplexity with five greedy-tail steps.
In briefNot All Is Lost: Fixing Lossy User Preference States in Personalization Encoders
SummaryThe paper observes that personalization encoders compress evolving interaction histories into preference states, and a task head using only that state can miss useful evidence still present in the frozen encoder's per-timestep cached representations. The authors propose REPAIR, which compares these cached representations with the current preference state in a compact learned space, selects useful patterns at specific timesteps, and adds their aggregate correction to the state. Training only REPAIR, with encoder and task head frozen, improves MRR and nDCG@10 across twelve recommendation hosts on four datasets.
Parthiv Chatterjee, Dhiraj Golhar, Ummesalma Diwan et al.
In briefIQS-BO: In-Context Query Selection for Bayesian Optimisation
SummaryThis paper proposes IQS-BO, a Prior-data Fitted Network that learns which query to evaluate next through supervised learning on synthetic priors, rather than refitting a surrogate or numerically maximising an acquisition function. In one forward pass, it predicts the probability that each candidate maximises the objective. The authors report that it matches or outperforms Gaussian process Bayesian optimisation and existing in-context methods on synthetic and real-world benchmarks. They also introduce a mixture prior for pre-training that includes warped, narrow-optima, and plateau functions.
In briefInstant Feedback Without Delay: Piloting a Generative AI Practice Tool in a Large Maths Class
SummaryThis paper reports a pilot in a large electrical engineering maths class where researchers built a generative AI practice platform. It gives students immediate, step-by-step feedback during self-directed practice. Human oversight is moved from live grading to checking solutions beforehand. The study examines how students used the tool, how they judged the value and reliability of its feedback, and which lessons might apply to other engineering subjects.
In briefPROMO: Preference-conditioned multi-objective RL for quadruped robots
SummaryThe paper presents PROMO, a method that lets a single quadrupedal locomotion policy take an operator's preference as a runtime input to balance command tracking, stability, and energy efficiency. In simulation, 67 of 100 sampled preferences yield non-dominated behaviors under exact Pareto dominance. The same policy transfers zero-shot to a Unitree Go2, where changing preferences alone reduces specific energy by up to 30.4%.
WhyConditions a single locomotion policy on runtime preferences and transfers zero-shot to a real Go2, backed by 100 simulated preferences and hardware results.
In briefScience Utopia? LLM-Agent Simulation of Academic Research Ecosystems
SummaryThis paper introduces SciUtopia, a persistent closed-loop simulation of academic research ecosystems driven by LLM agents, covering topic choice, collaboration, submission, peer review, resubmission, citation, funding, and attrition. Across 61 simulated worlds with over 40,000 researchers, the authors find that rejection-driven resubmission substantially increases reviewer burden beyond population growth. They also report that cautious exploration balances citation impact, career success, and topic diversity, and that resource inequality can emerge without detectable cumulative advantage from early funding wins.
WhyProposes a closed-loop LLM-agent framework simulating submission, review, citation, and funding across 61 worlds with counterfactual comparisons; conclusions depend on LLM simulation fidelity.
In briefAutomatically Sorting App Review Emotions with Large Language Models
SummaryThis paper studies how large language models can automatically assign multiple fine-grained emotions to mobile app reviews, based on a Plutchik-derived annotation framework with severe class imbalance. The authors compare fine-tuned encoders with zero- and few-shot prompting of decoder models, along with several imbalance mitigation strategies. The best few-shot decoder reaches a macro-F1 of 0.642, and the best encoder closes much of the gap through generative data augmentation and positive-weighted loss, at much lower inference latency. The experimental pipeline, synthetic corpora, and fine-tuned checkpoints are released.
In briefDeFA: dependency-based method for finding why LLM agents fail
SummaryDeFA is a framework that identifies the decisive error behind failures in LLM agent executions by analyzing dependencies between steps. It builds an event dependency graph across the trajectory and traces events that may violate task requirements to form a failure propagation graph. The method then pinpoints the key error, the responsible agent, and the error category. The paper reports high attribution accuracy on benchmarks and notes that its diagnoses improved downstream task accuracy when used for skill evolution.
In briefRevision-Aware Independent Agent Graphs for Dynamic Reasoning: a multi-agent method that picks the right document version before answering
SummaryThe paper studies dynamic task routing, where an event stream revises task bindings and the system must select the valid document version for each query before solving it. The authors convert six benchmarks (MMLU, MMLU-Pro, MedMCQA, MATH, GPQA, HumanEval) into 31,119 dynamic episodes with 373,428 queries. Their RIAG method caches solutions by document identity and conditionally audits and repairs them, reaching 54.24% joint accuracy at 0.62 calls per query.
In briefCorrect Answers, Corrupted States: Hidden Information Failures in AI Teamwork
SummaryThis paper argues that multi-agent AI systems can reach a correct decision while leaving behind a corrupted shared information state, which it calls an off-query failure. It introduces OffQuery, a benchmark that separately tests evidence verification, shared-state reconstruction, and task resolution in healthcare and disaster response settings. Results across GPT, Gemini, and Qwen models show task resolution is far stronger than state reliability. The authors also propose ReGround, which resolves conflicting evidence and reasons over a reconstructed trusted state, and report gains on all three capabilities across seven models.
In briefWhen the Judge Acts: Auditing Vision-Language Model Image Selection on Cultural Prompts
SummaryVision-language models increasingly serve as judges that choose the best of several generated images, shaping what users see. This paper audits these judges by their returned images rather than by score agreement with humans, comparing choices against unseen human ratings and random selection on 300 culturally situated prompts, with candidates reordered. A 4B-parameter judge barely beats random and trails a CLIP baseline, showing strong position bias. An 8B judge shows little position bias and outperforms CLIP, but agreement-based filtering mostly discards its good decisions.
In briefTesting How Well Japanese LLMs Handle IME and Typing Errors
SummaryThis paper evaluates how robust Japanese large language models are to realistic Japanese-specific typographical errors. The authors define five typo categories, including character transposition, character replacement, and IME-based conversion, and apply them to three Japanese benchmark datasets. Eleven Japanese and multilingual models were tested. Character transposition and replacement consistently reduced accuracy, while IME, full-width, and homophone conversion had relatively limited effects.
In briefMixture-Trained Merging: Training Mergeable Branches for Unified Multi-Objective Models
SummaryThe paper proposes Mixture-Trained Merging (MTM), which trains each branch on an objective-biased data mixture rather than a single objective so that branches remain compatible when merged. MTM uses merged-model evaluations as a low-cost signal to select branch mixtures and applies qNEHVI-based multi-objective Bayesian optimization iteratively. Experiments across code, mathematics, instruction following, and think/non-think control show MTM outperforms naive merging and preserves behavioral separation where single-objective merging collapses.
In briefLearning to Ask: Budget-Constrained Information Acquisition for Small-Large Model Collaboration
SummaryThis paper frames collaboration between a small language model (SLM) and a large language model (LLM) as an information acquisition problem under an API budget, rather than as computation allocation. The SLM remains the primary reasoner and queries a black-box LLM advisor only when needed, using targeted questions. The authors propose a three-stage RLVR framework that learns when to call the advisor and how to formulate useful queries, and report improved performance-cost tradeoffs on mathematical reasoning and coding tasks.
In briefHarness Annealing: Teaching Agents to Act with Less External Control
SummaryThis paper studies how language agents can internalize control decisions, such as tracking state, organizing workflows, and verifying answers, that are normally supplied by external harnesses. The authors propose HARNESS ANNEALING TRAINING (HAT), which combines explicit control supervision with a curriculum of teacher trajectories collected under progressively weaker harnesses. In experiments with 9B and 35B models on SWE-QA and SWE-QA-Pro, some annealed checkpoints using only tools score close to their starting checkpoints deployed with the full harness.
In briefASCRIBE: A Reasoning Framework for Generating Thai Clinical SOAP Notes
SummaryThe paper proposes ASCRIBE, a reasoning framework that assigns a clinical-significance level to each extracted fact from a physician-patient conversation before summarizing it into a SOAP note. The authors also release ThaiClinicBench, a de-identified benchmark of real Thai clinical encounters, and a synthetic training corpus derived from real clinical notes. According to the paper, ASCRIBE as a prompt outperforms chain-of-thought prompting on physician-aligned LLM-judge metrics, and as a GRPO reward it lets a Gemma-4-E4B model trained only on synthetic data match Gemini 3.1 Pro in factual precision and exceed it in completeness.
Tarm Kalavantavanich, Teerawut Ponarchar, Pattaramanee Arsomngern et al.
In briefJudgement in the Age of Jev: Cheap Machine Evaluation and Organisational Judgement
SummaryThis paper argues that generative AI has made producing plausible artefacts cheap, making evaluation and discernment the new constraints. Using TypeSafe AI's Jev model as a provocation, it proposes a conditional Jevons hypothesis: if the total marginal cost of usable machine evaluation falls enough, organisations may consume more of it. The article integrates rebound economics with research on authority, reliance and judgement to examine when cheap machine evaluation substitutes for human evaluative work and how it shapes the grounds for consequential organisational commitments.
In briefHHR: A Hierarchical Hash Retrieval Method for Faster LLM Generation
SummaryThis paper proposes Hierarchical Hash Retrieval (HHR), a coarse-to-fine framework that speeds up long-context inference in large language models. It combines Geometry-Aware Key Routing and Learned Hash Projection to reduce false positives and false negatives caused by binary hash encoding. The authors report a 1.10-point average gain on LongBench and, at 128K context, up to 3.30x decoding and 2.83x end-to-end speedup for Llama-3.1-8B-Instruct.
WhyProposes a coarse-to-fine hash retrieval fixing Hamming-attention mismatch, with LongBench gains and measured 128K decoding and end-to-end speedups.
In briefA compact explicit 4D model for representing moving scenes from video.
SummaryThis paper presents Sparc4D, a feed-forward autoencoder that compresses a monocular video with known cameras into a sparse 4D scene state. Static features are shared across the clip, and time-varying features are stored in spatially anchored temporal slots, using about 5x fewer values than per-frame storage. On a 32-frame MultiCamVideo clip, it reaches 21.97 dB on the first-32 protocol versus 20.40 dB for MoVieS, and it transfers to DyCheck and Neu3D without fine-tuning on real data.
In briefSupervising Success Criteria: Auxiliary Losses for Latent World-Model Planning
SummaryLatent world models plan by scoring candidate action sequences using distances in latent space, yet task success depends on physical quantities. The authors find that in all four models they examined, end-effector position is encoded with error exceeding the success criterion's tolerance. They propose an auxiliary loss that trains the model to regress these success-criterion quantities, using a linear head discarded after training so test-time models are unchanged. This loss alone raises success rates on PushT and cube by 3.5% and 3.4% absolute, both statistically significant.
In briefHaving an LLM write anomaly detectors for time series: compact, interpretable, automatically discovered
SummaryThe paper uses a large language model as the author of an anomaly detector rather than the detector itself. An autonomous loop repeatedly edits a short NumPy program under a leakage-free objective and keeps the best-scoring version. This yields two compact detectors, one univariate and one multivariate, that describe short windows by local spectral features and compare them to the training distribution with a covariance-aware distance. On the TSB-AD benchmark they lead strong classical, deep, and foundation-model baselines, including Time-RCD, across metrics. They train no network and use no GPU, and the multivariate detector is faster than every similarly performing baseline.
In briefHow Reputation, Strategy, and Emotion Shape Cooperation in Reasoning and Non-Reasoning AI Models
SummaryThis study uses the iterated prisoner's dilemma to test whether frontier generative AI models cooperate, varying counterpart reputation, strategy (extortion vs. generosity), and facial emotional signals. In a first study with non-reasoning models, all three factors systematically shaped cooperation, though models weighted them differently. A second study with reasoning models found greater reliance on strategy and reputation, a more conditional role for emotion, and model-specific end-game behavior that reveals differing exploitability. The authors argue for standardized cooperation benchmarks to guide responsible deployment in interactive settings.
In briefAGO AI Quality Gate: Evidence-First Release Decisions for RAG Systems
SummaryThis paper presents AGO AI Quality Gate, a framework that helps teams decide whether to promote, revise, or block a retrieval-augmented generation (RAG) system version. It combines a four-state decision model, layered scoring, a stratified beta-binomial gate for regression risk, and mandatory meta-evaluation of LLM judges. Tested on the public RAGBench benchmark, a low-cost judge performed barely above chance at detecting non-adherent answers, while a stronger judge performed substantially better but still varied across domains.
Giulio Zeloni, Enrico Lo Conte, Salvatore Rionero et al.
In briefAutoGUIWorld: Using Image Generators as Visual World Models to Train GUI Agents (a framework that generates GUI interaction trajectories without running real software)
SummaryThis paper presents AutoGUIWorld, a framework that combines an image generator's visual priors with a planner's task knowledge to synthesize GUI interaction trajectories without deploying the underlying software. The planner specifies atomic actions and their visual effects, and the image generator iteratively edits screenshots to produce subsequent observations. After grounding and quality filtering, the authors obtained 79,266 spatially annotated step-level training samples across Ubuntu, Windows, macOS, and Chrome. Fine-tuning Qwen3.5-35B-A3B on these trajectories raised its OSWorld mean task score from 33.0% to 40.8% and its ScienceBoard task success rate from 14.0% to 32.2%.
WhyNovel use of image generators to synthesize GUI trajectories without running software; clear gains on OSWorld and ScienceBoard support the claims.
In briefEgoFound3R: end-to-end hand reconstruction from first-person video in world coordinates, with point-wise interaction attributes
SummaryEgoFound3R is a unified end-to-end model that estimates hand geometry in world coordinates on a metric scale shared with the scene. It also predicts point-wise interaction attributes such as visibility, contact, and distance in a single pass. The model combines structured hand prompts, an explicit hand representation, and a shared-parameter multi-rate design. On OakInk-v2, TACO, and HOI4D it reduces MPJPE by 43.2%, 22.4%, and 11.6% over previous methods, with about 6x higher throughput.
In briefDependency-Aware Reward Shaping for Agentic Reinforcement Learning (DARS)
SummaryThe paper addresses the problem that terminal success or failure rewards in reinforcement learning for language models give little guidance about which steps contributed. It proposes DARS, which represents task progress as predicates linked by prerequisite relations and assigns step-level credit over this dependency graph based on annotations of verified, invalidated, or repaired predicates. The method is reported to work with existing training approaches across five task families and models from 1.5B to 8B parameters, with improvements on several benchmarks.
In briefSeparating Foreground and Hidden Scene 3D Reconstruction from Mixed Single-Photon LiDAR Echoes
SummaryThe paper addresses mixed LiDAR returns produced when partially transmissive screens or covers sit in front of a hidden scene. It proposes a state-aware framework that estimates each ray's echo state and uses it to supervise a two-head neural field for waveform reconstruction and geometry localization. The authors also release a real paired single-photon LiDAR occlusion dataset. Experiments on this data report better hidden scene depth and point-cloud accuracy than baselines.
In briefAssessing surgical skill from microscope stereo videos using combined RGB and depth information
SummaryThis paper proposes a semantic RGB-Depth framework for assessing surgical skill from microscopic stereo videos. It fuses sparse stereo depth with dense monocular depth into a geometric scene representation and combines it with instrument- and anatomy-specific RGB streams using a hierarchical attention architecture. Evaluated on 33 ex vivo transoral microlaryngeal procedures by six surgeons with leave-one-surgeon-out validation, the model reached an F1 score of 0.938 for skill-level classification, above the semantic RGB-only (0.696) and semantic depth-only (0.929) variants.
Jecia Z. Y. Mao, Sue M. Cho, Francis X. Creighton et al.
In briefAutoregressive Drillhole Modelling Under Distribution Shift: A Benchmark for Layer-by-Layer Lithology Prediction
SummaryThe paper introduces DrillBench, a benchmark of 49,671 Western Australian drillholes for predicting the next rock layer and generating stratigraphy autoregressively under increasing distribution shift. It finds that spatial and geochemical conditioning helps locally but degrades sharply under stronger shift, while lithology-sequence autoregressive models transfer more robustly. The authors combine large-scale pretraining on historical drillholes with spatial retrieval of neighbouring lithology, which improves generalisation under spatial and cross-province shift.
In briefiSEE: Keeping Track of Hidden Objects Through Self-Supervision
SummaryThis paper proposes iSEE, a framework that tracks an object's identity and position through occlusion in video without using any labels. It detects occlusion by comparing an object slot's attention with its past, separates appearance from position, and trains a position tracker using only where the object reappears. On the LA-CATER static dataset, the authors report that iSEE returns reappearing objects to their original slot after 86% of occlusions, compared with 32% for SlotContrast.
Pramish Paudel, Ajad Chhatkuli, Luc Van Gool et al.
In briefSelecting a small labeled subset using entropic optimal transport
SummaryThis paper studies low-budget active learning, where only a small coreset of points can be labeled and a model must reach high accuracy using just that selection. The authors use features from a pretrained self-supervised model and choose the coreset with the Sinkhorn divergence, an entropic optimal transport criterion, optimized by gradient methods and swap-based local search. Experiments on image benchmarks and medical datasets report that the method outperforms state-of-the-art heuristics in low-budget settings.
Rim Hajal, Mathieu Besan\c{c}on, J\'er\^ome Malick
In briefFederated Agent Optimization: Letting distributed AI agents improve together without sharing raw experience
SummaryThe paper formulates Federated Agent Optimization (FAO), a problem in which distributed LLM agents improve collaboratively through controlled information exchange while keeping raw data, full trajectories, and private knowledge local. It defines FAO as a multi-objective problem balancing agent utility, privacy leakage, and communication cost, and organizes the optimization space across policy, memory, tool use, reward, and structured knowledge and skills. The authors discuss how private experience can be abstracted, protected, aggregated, and adapted into transferable capabilities, and they outline key challenges and future research directions.
In briefGenerating counterfactual outcomes with flow matching, with end-to-end rate guarantees tied to the source–target coupling
SummaryThe paper proposes a flow-matching method for sampling counterfactual outcomes from observational data, combining a sample-split doubly robust objective with a learned coupling between source and target outcomes and a score-corrected stochastic sampler. It derives a KL error bound for Euler discretization controlled by moments of the source–target displacement under the chosen coupling, along with finite-sample guarantees for the learned velocity and score fields. Experiments on synthetic and semi-synthetic image benchmarks indicate that the stochastic sampler can outperform the deterministic ODE sampler at finite discretization budgets.
In briefFlashBack: Deciding When Streaming Vision-Language Models Should Recall Memory
SummaryStreaming vision-language models must handle continuously growing video under limited compute, balancing real-time perception against long-term memory. This paper introduces FlashBack, a training-free framework that first judges whether a query needs historical evidence, then either keeps the native inference path or uses a separate recall path combining recent context with retrieved memory. Evaluated on OVO-Bench and StreamingBench, it improves several long-horizon and memory-dependent tasks while largely preserving real-time perception.
In briefA color-independent method for segmenting words in handwritten Bangla text images.
SummaryThis paper addresses word segmentation in handwritten Bangla text images, a crucial first step for optical character recognition (OCR). The method is designed to work on smartphone-captured images regardless of paper and ink color or type. A custom dataset including shadow interference and other possible obstacles was built for evaluation. On 7,374 words, the system generated 7,278 bounding boxes with 90.60% recall, 91.80% precision, and a 91.20% F1-score.
In briefWhen Exercise-Specific Joint Selection Helps: Auditing Evaluation and Control Design
SummaryThis paper examines whether gains from exercise-specific joint selection in skeleton-based correctness classification are reliable. The authors audit 1,057 repetitions from ten REHAB24-6 subjects, separating the effects of evaluation aggregation, joint subset structure, and temporal representation. They find the reported gain depends heavily on the evaluation estimand and requires structurally matched random controls. The study is exploratory and does not claim a new algorithm or clinical benefit.
In briefFully online decentralized learning for stochastic games with unknown independent chains
SummaryThis paper studies stochastic games where players' controlled state chains evolve independently and transition kernels are unknown, so each player sees only its own state and realized payoffs. The authors propose a fully online, decentralized, uncoordinated mirror-descent algorithm operating on occupancy measures, using one sample per time step. Under uniform-ergodicity and finite-coverage assumptions, the time-averaged regret decays at the O(T^{-1/2}) rate, depending on local state cover times rather than the joint state space.
In briefSkeleton-and-Strategy Prompting: Training-Free Help for Vision-Language Models Understanding Negation
SummaryVision-Language Models often answer incorrectly when visual questions contain negation. The paper proposes SSP (Skeleton-and-Strategy Prompting), a training-free in-context method that abstracts a question into a structural skeleton, retrieves similar-skeleton questions, and has the model summarize a one-sentence answering strategy that is prepended to the test input. According to the paper, SSP achieves state-of-the-art results on negation-focused VQA benchmarks while remaining computationally efficient.
In briefToken attribution across denoising steps to show how diffusion language models generate text
SummaryThis paper introduces Diffusion Layer Integrated Gradients (DLIG), a token attribution method for diffusion language models that extends Integrated Gradients to any layer and denoising step. It tracks how the model progressively commits to a self-generated or fixed completion for an input prompt. The authors show that DLIG corresponds to the IG axioms of completeness, implementation invariance, linearity, and symmetry preservation. The method is demonstrated on word-sense disambiguation, multi-hop graph reasoning, and sentence infilling.
In briefRethinking information bottleneck: splitting label-related structure from within-condition variation
SummaryThe paper argues that standard information bottleneck regularization uses one scalar, I(Z;X), which mixes label-relevant structure with residual within-condition variation, so nuisance information can persist in representations. The authors propose a dual-bottleneck formulation with a standard KL term for global capacity and a conditional KL term for within-condition information, and show the conditional term decomposes exactly into a within-condition information term and a prior-mismatch term. Experiments report the clearest gains in low-data classification and consistent improvements on dense prediction benchmarks.
In briefCAGE-NAS: A certified method for growing neural networks efficiently by function-space descent
SummaryCAGE-NAS decides in function space when a neural network should expand. It keeps the architecture fixed while a certified functional gradient descent step is admissible, and applies a function-preserving expansion when the check fails. In a controlled setting with exact certification, the resulting architectures reach above the 99.8th percentile by held-out RMSE among all admissible alternatives within the same parameter budget.
In briefLearning Rate Transfer Across Hybrid Transformer-SSM Model Architectures
SummaryThis paper studies how learning rates scale in hybrid models that combine Transformer and State-Space Model blocks. Using simplified-ZOH Mamba at fixed state size, the authors find that the original μP recipe keeps the learning rate transfer gap near zero across widths, depths, and billion-parameter scales, even though these models fail the standard coordinate check. They attribute this to a global update-to-weight invariance and AdamW's local per-parameter balance, and report that the effect also holds for the production model Nemotron-H.
In briefHeadEdit: Adjusting Language Model Behavior via the Frozen Unembedding Matrix
SummaryHeadEdit is a gradient-free method that calibrates language model behavior through the unembedding matrix, without updating parameters. It extracts a low-rank behavioral subspace from paired completions and uses each prompt's coordinates in that subspace to generate a vocabulary-wide correction. The authors report improvements across all nine experimental settings spanning three tasks and three model families, with negligible inference overhead and no systematic loss of general capabilities.
In briefDetect, Explain, Interpret: A Benchmark for Time Series Anomaly Detection and Explanation
SummaryThis paper introduces SHAD, a fully annotated benchmark of 215 multivariate, high-dimensional time series collected from real-world distributed cloud storage systems operated by Scality. The dataset includes contextual information and covers three families of anomalies with varying severity. The authors evaluate baseline methods across detection, explainability (whether per-dimension contributions yield accurate anomaly attributions), and interpretability (whether frozen LLMs can localize and interpret anomalies).
Roberto Stanzione, Jules Barbe, Magali Parrino et al.
In briefCineMR: A tool-using AI model for quantitative heart MRI analysis
SummaryThis paper presents CineMR, a vision-language model that calls cardiac image-analysis tools to perform quantitative assessment of cine cardiac MRI. The authors also build a multi-cohort visual question answering benchmark covering metric extraction, multiclass diagnosis, and differential diagnosis. CineMR reaches 35.9% pass@1, well above the baseline models, and performance drops to 27.9% when tools are removed, indicating that reliable tool use matters for quantitative reasoning.
In briefLayer Bypass Responses Across Pretraining Checkpoints: Depth Order Persists While Magnitudes Shift
SummaryThis paper examines how single-layer identity bypass responses in language models change across pretraining checkpoints. Using five released training trajectories and 11 model-domain combinations, the authors find that depth ordering of responses is retained while their magnitudes redistribute over training. They conclude that layer sensitivity is structured but not static, so single-checkpoint intervention results should be read in light of how the perturbation pathway evolves during training.
In briefPhysicsLENS: Testing whether video generators respect hidden physical properties
SummaryThis paper introduces PhysicsLENS, a dataset and benchmark for checking whether generated robot videos follow physical properties such as weight, friction, or viscosity. It uses matched scenario pairs with the same starting frame and task but different stated physics, covering seven domains including collision, gravity, and fluid. Across four video generation models, over 400 human-annotated labels were collected. Results show that plausible-looking videos often ignore the stated property (34 of 47), and stating the property lowers plausibility only slightly and not significantly.
In briefFAULT: Using Self-Diagnosis to Assign Credit in Self-Evolving Agent Reinforcement Learning
SummaryThis paper addresses credit-assignment problems in agentic reinforcement learning for language model agents, which arise from relying only on terminal outcome rewards in multi-step tasks. It proposes FAULT, a method that turns self-diagnosed errors into explicit step-level credit anchored by terminal outcomes, checking diagnostic evidence and learning relative error costs from task results. On ALFWorld, FAULT reaches 95% signal coverage, compared with 41% for GRPO and 72% for GiGPO, and better localizes credit to specific error steps. It delivers strong gains on long-horizon ALFWorld and WebShop tasks while staying competitive on short-horizon Search-based QA.
In briefLooping Beyond Twice: A Scalable Recipe for Looped MoE (LOOM)
SummaryThis paper studies looped Transformers built from mixture-of-experts (MoE) blocks, where repeated loops add depth without adding parameters. The authors find that gains flatten or reverse after about two loops due to growing hidden-state variance and routers repeatedly choosing the same experts. They propose LOOM, which scales residual updates, re-injects the input embedding each loop, and uses per-loop routers and a Looping Residual. Experiments on 100M–1.7B models report stable scaling to 9–12 loops with lower perplexity and higher zero-shot accuracy than non-looped baselines.
In briefBanglaDial-Abuse: A dataset for identifying regional dialects in abusive Bangla text
SummaryThis paper introduces BanglaDial-Abuse, a balanced Bengali-script dataset of 1,000 sentences for identifying regional dialects in abusive Bangla text. It covers four varieties (Standard Bangla, Chattagram, Sylhet, Barishal) with 250 samples each. The corpus was built with a corpus-grounded synthetic procedure that preserves the hostile meaning while incorporating regional variation. The authors state it is a research and prototyping resource, not a native-speaker-validated gold standard.
In briefWhen Deletion Ordering Is Tractable: Update Dynamics to Permutation Structure
SummaryThe paper studies how the order of processing pending deletion requests, one by one under a fixed policy, affects the final trained model. It identifies two structures that simplify the problem: position additivity, which reduces optimization to assignment (or sorting when positional profiles are shared), and suffix localization, which removes dependence on distant earlier requests. Under shared affine updates, the authors characterize the quadratic interactions that break additivity and show that suffix-conditioned assignment improves the approximation rate from O(p^L) to O(p^(2L)). Experiments show these structures can be recovered from executed objectives and predict unseen orders.
In briefOptimusMesh: compact mesh generation from point clouds using sparse latent pivots
SummaryOptimusMesh is a framework that generates compact triangle meshes directly from point clouds. It compresses 2,048 oriented input points into 16 sparse latent pivots and uses a two-stage autoregressive process that generates vertices first and then faces. Compared with the evaluated recent autoregressive methods, it produces 25.7% to 94.1% fewer faces while keeping competitive geometric fidelity and distributional quality.
In briefParameter-efficient robust adaptation of tabular foundation models under group shifts
SummaryThe paper notes that tabular foundation models can perform poorly on underrepresented groups when group proportions change between training and deployment. The authors propose DR-TFM, which adapts only a query scaling network without needing true group labels, keeping other parameters fixed. For TabPFN-3, adaptation updates just 0.016% of parameters. Across five tabular benchmarks, DR-TFM achieves substantially higher average worst-group accuracy while keeping mean group accuracy competitive.
In briefStress tests and adaptive routing for chest X-ray vision-language models
SummaryThis paper builds a reliability stress test for AI models that interpret chest X-rays, using two balanced datasets and comparing three medical vision-language models. The models are tested across three prompt styles and two workflows, giving 36 configurations. The authors find that exact-match accuracy alone can overstate conservative models that default to "Normal" predictions, and that reliability depends strongly on model family and scale. They propose a decision-time routing framework that uses multi-agent inference only when it helps, improving the cost-quality trade-off.
In briefReSolve: Reusing Intermediate Reasoning via Selective Model Review
SummaryThis paper introduces ReSolve, a training-free inference method that reuses existing candidate reasoning. When sampled solutions disagree or lack a parseable answer, a model examines the earlier derivations and adds its new solution to a bounded loop. On 130 competition-math problems, ReSolve answered more questions correctly than voting over the same four candidates and used fewer tokens than eight-sample self-consistency. The authors argue candidate reasoning is reusable computation, but they do not establish an accuracy advantage over extra sampling or a distinct benefit from specialized route instructions.
In briefDo Multilingual Encoders Produce Language-Consistent Semantic IDs?
SummaryThis paper examines whether multilingual encoders give different-language versions of the same product consistent semantic IDs for generative retrieval. Using Amazon ESCI listings in English, Spanish, and Japanese, the authors find that multilingual E5 places translations measurably apart from their English sources, with a Japanese translation keeping the first code only 7.7% of the time. Balancing the quantizer fitting data makes codebook use more uniform but further reduces cross-lingual prefix agreement. The authors conclude that multilingual exposure and balanced codebook use alone do not guarantee language-consistent semantic IDs.
In briefAuditing how concurrent LLM agent actions get settled: order, progress, and replay
SummaryThe paper studies how to arbitrate concurrent actions by LLM agents when each proposed action is individually valid. The authors implement a typed snapshot-settlement contract and test order sensitivity, useful progress, and replay consistency. Five settlement policies were evaluated across 28,800 permutation trials and 2,160 scripted multistep episodes.
In briefRSNA Intracranial Aneurysm Dataset: A Public, Multi-Institutional Imaging Resource for AI Aneurysm Detection
SummaryThis paper presents the RSNA Intracranial Aneurysm (RSNA-ICA) Dataset, curated with several neuroradiology societies for the 2025 RSNA Intracranial Aneurysm Detection Challenge. It contains 7202 CTA, MRA, and MRI series from 4278 adult patients across 21 institutions in 12 countries. Experts annotated aneurysm presence and location, and 178 series include three-dimensional vascular segmentations. Of these series, 5041 are publicly available through MIRA for noncommercial research use.
Maria Correia de Verdier, Rachit Saluja, Jason Sho et al.
In briefConfidence-gated cloud-edge triage for chest X-rays using variational risk minimization
SummaryEmergency chest X-ray triage faces a modality gap because reports arrive after triage decisions, while multimodal models need image-text inputs. The paper proposes Variational Risk Minimization (VRM), a distillation framework that treats LVLM-generated report variants as Monte Carlo samples of latent clinical interpretations. Under matched encoder families, VRM outperforms direct fine-tuning baselines and improves calibration. A compact edge student with a confidence-gated cascade reaches AUC 0.941 at 103ms average latency, with 20.3% cloud escalation.
In briefRecognizing Handwritten Bangla Words with Deep Object Detection Models
SummaryThis study uses deep learning to recognize words in handwritten Bangla word images, comparing SSD with MobileNetV2, Faster R-CNN with InceptionResNetV2, and an ensemble of both, with a modified Non-Maximum Suppression. A custom dataset of 9,841 handwritten Bangla word images was compiled. The ensemble model performed best, reaching an F1-score of 92.61% and correctly recognizing 96.12% of words to some extent.
In briefDoes Scaling RL Really Need More Training? A Method to Reuse Completed RL Histories
SummaryThe paper proposes policy-space scaling, which improves a reasoning policy by reusing checkpoints saved during a completed reinforcement learning run, without extending training or increasing per-query inference compute. The method, SURGE, fuses a high-accuracy anchor checkpoint with a donor checkpoint that produces shorter responses, using spectral decomposition of their updates. Evaluated on two 1.5B mathematical-reasoning histories and one 7B coding history, the fused model improves benchmark-average accuracy over both inputs while using fewer reasoning tokens than the anchor.
In briefUsing language models inside economic simulators to generate and test policy
SummaryThe paper places an instruction-tuned language model inside six Snowdrop-backed DSGE economic simulators, where at each turn it observes the economy, selects a bounded policy action, and receives the next simulated state and an economic reward. PPO is the primary method and GRPO a critic-free baseline. The goal is to judge policy actions by their simulated economic consequences rather than by how plausible they sound.
In briefA new tokenizer training and encoding method that yields shorter token sequences for LLMs
SummaryThis paper proposes CNF, a tokenizer training method that builds a vocabulary by counting valid substrings and then filtering them by actual usage, and MCE, an encoding algorithm that globally minimizes a segmentation cost. Across six text categories and two vocabulary-size groups, CNF-MCE compresses text better than the evaluated BPE tokenizers. Models trained at 1.8B and 8B scales achieve average performance comparable to BPE-based models across 11 benchmarks.
In briefSharing hidden-layer activations to speed up federated learning
SummaryThis paper addresses client drift, a key obstacle in federated learning caused by differences in data across clients. It proposes sharing a small amount of hidden-layer activations to reduce this heterogeneity. The authors report that the method improves training efficiency while preserving convergence guarantees and data privacy. Under a fixed round budget, it achieves higher accuracy than FedProx, SCAFFOLD, FedPVR, FedProto, and SplitFed without excessive communication overhead.
Seungjun Lee, Ensieh Khazaei, Dimitrios Hatzinakos et al.
In briefCortexBridge: Cortical Alignment Lets EEG Foundation Models Handle Different Electrode Layouts
SummaryEEG foundation models are usually pretrained on a fixed channel set or a few electrode layouts, which makes transfer hard when layouts change. The authors propose CortexBridge, a lightweight adapter that combines EEG features with electrode and atlas coordinates to map arbitrary montages into a shared cortical latent space. Tested with three frozen foundation models on five MOABB brain-computer interface datasets, it improved 13 of 15 evaluations, with average balanced accuracy gains of 0.80% (EEGPT), 0.70% (LaBraM), and 3.26% (CBraMod).
In briefAbsorbEvo: An agent framework for autonomously designing microwave absorbers from plain-language goals
SummaryThe paper presents AbsorbEvo, an agentic framework that turns natural-language performance goals into microwave absorber designs verified by full-wave simulation. A large language model proposes parameter adjustments, a low-cost predictive model ranks candidates, and only top designs are simulated. On held-out AbsorbBench-36 tasks, it reached a 79.17% task success rate, compared with 25.00% for a generic agent and 12.50% for random search.
In briefMadeleine: learning involuntary recall for chatbot memory from simulated lives
SummaryLong-term conversational assistants must recall the right memory at the right time, even when it is not similar to what the user is currently saying. Madeleine learns these associations offline by training a query encoder on cue-trigger pairs from LLM-simulated lives, adding a learned residual on top of frozen similarity. At inference it makes no LLM calls and can replace the query encoder in any vector memory. On LoCoMo-Plus, it reaches 66.6 when plugged into HyperMem and matches HyperMem's released score alone (52.4 vs. 52.9) with zero LLM calls.
In briefStepCAD: Generating CAD Code from Meshes Using an LLM Policy and Guided Search
SummaryThis paper presents StepCAD, a method that recovers executable CAD programs from 3D meshes by combining a state-conditioned CAD policy with geometry-guided tree search. The authors also release ARCADE-1.5M, a dataset of 1.5 million executable CAD programs covering diverse operations. According to the paper, StepCAD reaches state-of-the-art reconstruction accuracy across several benchmarks, with up to 87.2% relative IoU improvement over the strongest baseline.
In briefStandardising environmental impact metrics for AI research, beyond state-of-the-art accuracy
SummaryThis paper notes that the environmental footprint of large language models is growing, while the machine learning community lacks standard carbon accounting practices. The authors automatically reviewed the 5,285 papers accepted to NeurIPS 2025 and found that environmental impact reporting is nearly absent. They propose standardised sustainability metrics for training efficiency, simple heuristics for estimating LLM inference carbon cost, and carbonbenchmark, a software tool for tracking emissions. They also formalise the 'Smallest Model that Achieves the Job' (SMAJ) framework.
In briefAffine-aligned atlas method for building canonical Gaussians in video representation
SummaryThis paper proposes an affine-atlas canonical Gaussian representation for video reconstruction. Frame-wise affine transforms first absorb global motion such as camera movement, and canonical Gaussians are then built in a larger affine-aligned atlas space so they match individual frames better. The change affects only the canonical construction stage and can be added to existing canonical-Gaussian methods with negligible extra parameters. Experiments show improved reconstruction quality, especially on sequences with large camera motion.
Masaya Takabe, Hiroshi Watanabe, Sujun Hong et al.
In briefHow much language models gain from extra test-time computation, and at what cost
SummaryThis paper introduces SELF-POT, a benchmark that measures how much extra test-time computation improves language models across competition math, competitive programming, and agentic workflows. Under a unified budget, it compares direct inference with parallel sampling and self-revision, charging every model call, including selection and critique, in dollars. The results show that returns depend on the domain, the selection rule, and how failures are handled.
In briefAgSpec: A Retrieval-Based Speculative Decoding Framework for Coding Agent Pipelines
SummaryThis paper presents AgSpec, a framework that improves retrieval-based speculative decoding for coding agents by retrieving draft text from session, workspace, and global corpora. It also sets per-agent draft-length caps and adjusts draft length online using verification feedback. On two repository-level multi-agent coding benchmarks, AgSpec outperforms five retrieval-based drafters and EAGLE-3 in most settings, with throughput gains up to 4.37x at batch size 1 and 4.76x at batch size 16.
In briefMisconceptDialogue: How Evaluation Criteria Change Measured Success in LLM Math Tutoring
SummaryThe paper introduces MisconceptDialogue, a framework in which five open-weight LLM tutors (7B-14B) interact with LLM-simulated students over more than 5,000 dialogues to detect errors, diagnose misconceptions, and remediate them. Exact-match scoring reports a 42.1% remediation success rate, while a human-calibrated LLM judge reports 81.3% on the same dialogues, a 39.2-point gap. The tutors' scores barely change whether they receive the correct or a deliberately swapped diagnosis. The authors argue that multi-criterion scoring with multi-judge convergence and human anchoring should become standard for evaluating LLM tutoring.
In briefMASkillBlender: Decentralized whole-body coordination for multiple humanoid robots that move and manipulate objects together
SummaryThe paper presents MASkillBlender, a multi-agent reinforcement learning framework that lets multiple humanoid robots coordinate their whole-body movement and manipulation in a decentralized way. It learns a shared high-level policy that blends reusable pre-trained single-humanoid skills, using only task-level rewards rather than task-specific motion references. The authors also propose a permutation-based data augmentation for homogeneous robot teams and evaluate the framework in simulation on several coordination tasks across two humanoid embodiments.
In briefMVDG: Efficient Multi-view 3D Disambiguation for Real-World Images
SummaryThis paper addresses false matches between visually similar but distinct 3D surfaces, which hinder 3D reconstruction and visual localization. It proposes MVDG, a multi-view disambiguation framework built on the VGGT 3D foundation model that jointly processes an arbitrary number of images. The authors fine-tune on a pseudo-pairwise training set derived from AerialMegaDepth and validate efficiently with a pseudo-pairwise test set. Experiments report comparable pairwise accuracy with improved SfM accuracy and inference speed over baselines.
In briefYouRA: a persistent-state architecture for research agents with traceable evidence
SummaryEnd-to-end research agents can write full papers, but their claims often differ from what the experiments actually showed. YouRA keeps research state, execution evidence, and failure history persistent across the research process, using three components: a verification state architecture, an independent controller, and stateful reflection. On a ten-task end-to-end subset of MLR-Bench, the paper reports that YouRA scores higher on overall performance than MLR-Agent and AI Scientist V2 across three matched backbones.|
In briefDataset Identity, Not Novelty: Why OOD Detection Gains Are Inflated
SummaryThe paper examines post-hoc out-of-distribution detectors that are fitted on the same OOD dataset used for evaluation, arguing that standard held-out validation can reward recognizing which dataset is being seen rather than true novelty. The authors measure this inflation by holding out the entire OOD dataset and test it across combiners on ImageNet and CIFAR-100 backbones. They report that most of the usually reported gain is attributable to dataset identity.（约 80 词）
In briefOrbitTAMP: Connecting Language Models to Spacecraft Rendezvous Task and Motion Planning
SummaryThis paper proposes a hierarchical framework for task-and-motion planning in spacecraft rendezvous and proximity operations. A pretrained language model turns a natural-language command into a partial mission specification, and planners fill in unspecified decisions within the admissible operational space. Trajectory optimization then produces a dynamically feasible trajectory. The authors report that this architecture improves intent recovery over direct language model generation.
Yuji Takubo, Daniele Gammelli, Marco Pavone et al.
In briefEgo2Act: a benchmark testing whether video models can generate goal-directed hand manipulation
SummaryEgo2Act is a benchmark of 2,640 egocentric videos covering 110 real-world everyday tasks, used to test whether video generation models can create a hand carrying out multi-step object manipulation toward a high-level goal from an initial scene image. The authors also propose Ego2ActJudge, a reference-free evaluation pipeline that aligns better with human consensus on task completion and physics plausibility than relevant baselines. The study finds that generated videos often skip or only partially execute steps, leaving the goal unfulfilled, and that models consistently struggle with fine-grained physical dynamics during complex manipulation and persistent world modeling.
Patrick Amadeus Irawan, Iskandar Muda Rizky Parlambang, Rava Maulana et al.
In briefRandomly reweighted NPMLE for Gaussian mixtures has polylogarithmic support size
SummaryThe paper studies the nonparametric maximum likelihood estimator (NPMLE) for Gaussian location mixtures, which can have many atoms. The authors add a vanishingly small random perturbation to the likelihood via Gamma weights, yielding a randomly reweighted NPMLE. They show that with high probability it is unique and has O{(log n/log log n)^d + log n} atoms in dimension d, while nearly maximizing the ordinary likelihood. Numerical illustrations report Hellinger risk and support size comparable to the ordinary NPMLE.
WhyRandom reweighting yields polylogarithmic NPMLE sparsity, improving the linear atom bound, with an effective-dimension proof and numerical checks.
In briefPrecision Over Scale: A Polish-Silesian Benchmark and Translation System
SummaryThis paper examines machine translation between Polish and Silesian, a dialect that standard benchmarks often fail to capture. The authors compare neural and rule-based systems and introduce SiLTT, a new Polish-Silesian test set, alongside the BOUQuET and FLORES benchmarks. The rule-based system performs most consistently on SiLTT and BOUQuET, while a TranslateGemma model fine-tuned on curated data beats strong neural baselines but does not surpass the rule-based system. The authors release SiLTT and their best neural model.
Grzegorz Kulik, Miko{\l}aj Pokrywka, Adam Jatowt et al.
In briefEvent-Guided Neural Video Compression: Using Event Streams to Cut RGB Video Bitrate
SummaryThis paper proposes ENVC, an event-guided neural video codec that uses event streams, which record brightness changes between frames and are shared by encoder and decoder, to improve RGB video compression. It forms an event-guided motion prior and codes only the remaining motion residual, and uses an event-conditioned predictor to supply multi-scale features for refining temporal context. Across six benchmarks, ENVC achieves average BD-rate savings of 39.13% with PSNR-RGB and 67.63% with LPIPS relative to DCMVC.
Jiyun Kong, Jungwoo Kim, Enes Eray Demirtas et al.
In briefValue-guided search to better train models on math reasoning
SummaryThis paper proposes APIVIS, a training framework for reinforcement learning with verifiable rewards that applies finite-budget Gumbel search to chunk-level math reasoning. It combines direct and searched responses within each rollout group so that search improvements yield informative relative rewards, and applies selective supervision to search-improved tokens. The authors provide theoretical guarantees on expected verifier reward and report experiments on mathematical reasoning benchmarks showing gains over competitive search-based methods.
In briefGLoC-EHR: Clinical reasoning from patient records that cites its evidence
SummaryThis paper introduces GLoC-EHR, a multimodal language model that reads structured electronic health records using a global memory of the whole patient trajectory and a local memory of selected events. It is trained to generate hospital-course summaries and to describe masked concepts, then to cite record evidence before answering, using rationale fine-tuning and GRPO. The authors report macro AUROC results on three MIMIC-IV outcome tasks against zero-shot LLMs, along with transfer results on EHRSHOT.
In briefExplains how to scale logic gate networks without wasted, redundant logic kernels
SummaryThis paper studies how to scale differentiable logic gate networks, efficient alternatives to conventional neural networks built only from logic gates, by increasing their width. The authors find that naive width scaling creates redundancy among logic kernels, causing performance to saturate. They propose a dynamic logic kernel framework that promotes specialization across kernel groups and uses input-dependent routing implemented with Boolean operations at inference, with an early-stage strategy targeting the first gate level where redundancy is strongest. Experiments show improved kernel utilization and diversity, yielding higher accuracy and better parameter efficiency.
In briefUsing Participation Trophies to Probe LLM Ability: Random-Reward RL as a Capability Probe
SummaryThis paper proposes random-reward reinforcement learning as a probe of a language model's reachability, meaning what further training can achieve from its current state under given constraints, not just its present performance. Two OLMo checkpoints with equal 3.5% accuracy on synthetic arithmetic reached 8.5% and 55% in their best correctness-rewarded RL runs. The authors report three training regimes across pre-training and mid-training, where random rewards become effective only in mid-training. They argue this helps distinguish whether a probe reveals model capability or learns the task itself.
In briefJoinGR: Learning to Traverse Join Graphs for Table Retrieval
SummaryThis paper introduces JoinGR, a table retrieval method that uses a database's join graph to find relevant tables for Text-to-SQL. It identifies tables not mentioned in the question by following join relationships from semantically similar anchor tables. JoinGR is competitive with strong baselines on BIRD and Spider, and substantially improves recall over dense retrieval and re-ranking on the multi-hop enterprise benchmark BEAVER.
Sandipan De, Abhijit Chakraborty, Sambaran Bandyopadhyay et al.
In briefKernelized Activation Steering: locally adaptive control of model activations via kernel methods
SummaryThis paper proposes Kernelized Activation Steering (KAS), which lifts activation steering into a reproducing kernel Hilbert space and computes steering scores purely through kernel evaluations. Unlike Difference-in-Means, which applies one fixed steering vector to all activations, KAS adjusts each activation according to its position relative to source and target reference sets. The authors state that DiM is recovered under a linear kernel, and richer kernels allow geometry-aware interventions. KAS is reported to outperform or match existing methods on tasks including LLM jailbreaking and image style control.
Laziz U. Abdullaev, Minh-Hieu Pham, Bach Do et al.
In briefHierGF: Hierarchical Gaussian fields for reconstructing 3D objects from very few images
SummaryThis paper addresses 3D reconstruction from sparse views, where only a few images of a scene are available. The proposed method, HierGF, uses a two-stage geometry-perception network to turn coarse 3D geometry and 2D generative priors into structured pseudo-supervision. It adds a learnable confidence network to favor cross-view consistent content and a geometrically consistent densification module to improve under-sampled regions.
In briefCapturing in-context learning with task operators: a method that extracts and reuses how models learn from examples
SummaryThis paper studies how language models learn from demonstrations in context without weight updates. The authors show that each attention head's output is an affine transformation of its context-masked version, with parameters stable across samples for a given task. They introduce Task Operator (TO), which applies this transformation as an analytically derived update to the attention output projection. TO achieves the best overall performance among prior methods on lexical, algorithmic, and reasoning tasks, narrowing the gap between zero-shot inference and in-context learning.
