In briefWhen Evidence Changes: Testing Memory Repair Versus Re-reading for Language-Model Agents
SummaryThis paper examines whether a language-model agent should repair its memory or re-read current evidence when the documents behind its derived facts are revoked or replaced. The authors evaluate caching, rebuilding, and graph-local repair against full and source-filtered re-reading on ICU medication and problem-list tasks, using two 7B models. Memory pipelines cost at least twice as much as full re-reading in held-out conditions, and source-filtered re-reading was cheapest. None of the four primary confirmatory tests in the replacement study reached statistical significance.
Wenhui Chu (University at Albany, State University of New York)
In briefAdaEva: Speeding Up LLM-Driven Algorithm Design with Adaptive Partial Evaluation
SummaryThis paper addresses the high cost of evaluating algorithms generated by large language models, which must normally be tested on an entire set of training instances. The authors propose AdaEva, a framework that evaluates candidates on progressively larger subsets of instances and removes unpromising ones as evidence builds up. It works without changing the underlying LLM-driven design procedure and needs no prior knowledge of instance difficulty. Experiments across several frameworks, LLM backbones, and optimization domains show better search efficiency and held-out generalization under matched evaluation budgets.
In briefProxy Confidence: Auditing Black-Box LLM Agents with a Surrogate Model's Probabilities
SummaryThe paper addresses deployed LLM agents that can silently issue wrong tool calls, where chat APIs hide token probabilities and self-stated confidence is weak. It proposes running a low-cost open-weight surrogate model in parallel to score the agent's proposed action from its log-probabilities, without accessing the agent's internals. On difficult coding tasks the method reaches AUROC 0.825, versus 0.598 for the actor's stated confidence. The authors also report gains in two deployment modes, a real-time review gate and confidence feedback to the agent.
WhyScores black-box agent tool calls via an open surrogate's log-probabilities with complementary readouts and a selection principle; validated across actors and benchmarks with task-success gains.
In briefREACT: A Framework for Physically and Chemically Consistent Reconstruction of Ocean pH
SummaryThis paper presents REACT, a framework that reconstructs global sea surface pH from sparse observations to help monitor ocean acidification. It transports dissolved inorganic carbon conservatively, corrects non-conservative carbon-cycle changes with a source module, and decodes the result into pH under carbonate equilibrium. On simulation data, REACT reduces pH NRMSE by 14.7% and chemical consistency error by 24.0% relative to the best baseline.
In briefReinforcement learning for a crystal generator, with validated gains and reward hacking
SummaryThe paper applies group-relative policy optimization to a crystal generative model built on stochastic interpolants and discrete flow matching, aligning it with general black-box reward functions. The reward raises the yield of metastable, unique and novel (mSUN) structures from 13.4% for the pretrained model to 45.5% for the reinforced model, per a community benchmark. The authors also report reward exploitation when atom-type transition likelihoods are reinforced directly, and a benchmark gap where single-element structures in different packings inflate the metric.
Pawan Prakash, Philipp H\"ollmer, Addis Fuhr et al.
In briefStreaming multi-track timeline control for 3D human motion generation
SummaryThis paper introduces TimelineControl, a method that lets a generated human motion respond to new text instructions while ongoing actions continue, such as answering a phone while walking. It uses interval-aware conditioning to keep instruction timing, and causal part-structured representations with part-aware denoising to coordinate concurrent actions across body regions. The authors also release TimelineMotion, a dataset with overlapping instruction intervals and body-part annotations. Experiments on TimelineMotion and MTT report better semantic alignment and temporal adherence than the evaluated streaming baselines.
Yangsong Zhang, Anujith Muraleedharan, Rikhat Akizhanov et al.
In briefTraining Numerical Intelligence via Auto-Diagnosis and Skill Discovery
SummaryThis paper introduces ADSD, a framework that first diagnoses why a numerical solver performs poorly, then uses that diagnosis to discover suitable numerical methods. The resulting knowledge is packaged as reusable solver skills. Across four domains (power flow, AC optimal power flow, stiff ODEs, and heterogeneous diffusion PDEs), ADSD improves accuracy, robustness, and efficiency. On GOC-500 power flow, mean solver error is reduced by nearly 71 times, and the gains transfer to unseen grid topologies and operating regimes.
In briefSimpler decoders, better encoders: geometric representation learning via novel view synthesis
SummaryThis paper studies how Novel View Synthesis (NVS) can help learn 3D-aware visual representations. The authors argue that poor results from existing encoder-based NVS methods stem from overly expressive decoders and low-level pixel-space targets rather than a lack of supervision. They propose SNAP, a self-supervised encoder-decoder transformer with a pose-conditioned local decoder and latent-space reconstruction, which is competitive across five tasks including visual localization, pose estimation, and robot manipulation.
Keerthi Kaashyap, Dennis Anthony, Akshay Krishnan et al.
In briefMoSE3: Learning Dense 6-DoF Motion for Every Pixel in World Space
SummaryMoSE3 is a feed-forward model that predicts dense SE(3) motion from monocular RGB video, giving a full 6-DoF rigid transform for every pixel in world space. It learns 3D point tracks and rigidity embeddings jointly, then recovers transforms by differentiably fitting them within soft rigid clusters. The authors also introduce Art-Kubric, a synthetic dataset with dense SE(3) and rigidity labels for articulated objects. According to the paper, MoSE3 achieves state-of-the-art SE(3) estimation and 3D point tracking accuracy and generalizes to real-world videos despite training only on synthetic data.
WhyFirst feed-forward dense world-space SE(3) prediction with a new synthetic dataset; state-of-the-art on several benchmarks, impact mainly within motion modeling.
In brief4DCodeBench: A Benchmark Testing Whether AI Agents Can Recreate Moving Scenes as Code
SummaryThis paper introduces 4DCodeBench, a benchmark for 4D inverse graphics in which agents reconstruct dynamic scenes from video as executable graphics programs. The authors built a test set of real-world videos and synthetic scenes covering phenomena such as deformation, fluid flow, and fracture. Benchmarking frontier models showed that strong static reconstruction does not yet reliably carry over to complex dynamics.
Ruihong Shen, \v{Z}iga Kova\v{c}i\v{c}, Peter Kulits et al.
In briefGOTT: Object-centric dexterous manipulation using a reusable cross-robot contact primitive
SummaryThis paper proposes GOTT, a reach-acquire-move framework built around one contact-acquisition primitive that can be reused across different robot embodiments. The hand first moves near a task-relevant contact region, then a closed-loop primitive establishes stable contact, and a pose-conditioned controller tracks the desired object motion. Experiments in simulation and the real world report robust contact across diverse objects and hand morphologies, including unseen ones, and higher end-to-end task success than open-loop grasp execution.
In briefWhat World Models Should Forget: Layered Retention for Continual Adaptation
SummaryThe paper argues that standard continual learning, which treats performance drops on past data as failure, does not fit world models, since environments change and outdated knowledge should be revised or discarded. The authors propose retention stratified by invariance timescale: physical laws and object permanence must never be revised, while instance-level facts should be updated as soon as the environment changes. They introduce differential retention, which reports invariant regression testing alongside revision latency without aggregating them into one score.
In briefRNADyn: A Benchmark for Generating and Understanding RNA Dynamics
SummaryRNA function depends on conformational changes that static structures do not fully capture. The authors introduce RNADynBench, a benchmark of 2585 quality-controlled 100-ns all-atom molecular dynamics trajectories with leakage-controlled data splits. They also present RNADynNet, a unified model that uses a shared backbone to generate trajectories and extract dynamics fingerprints from a single conformer, with physical grounding. On both test sets, the generated trajectories reach RMSF correlations of 0.875 and 0.766.
In briefEyeRobot 2.0: Precise two-handed robot manipulation using one stereo camera and active gaze, without wrist cameras
SummaryThis paper introduces EyeRobot 2.0, which uses a single stereo camera and active gaze to keep two viewpoints centered on a 3D fixation point, with more visual tokens given to image centers. A low-level gaze policy and a target selector are trained with reinforcement learning on real-world data. In 7 real and 6 simulated tasks, it outperformed passive stereo by 40% in real-world trials and 20% in simulation. It matched wrist-camera policies when wrist views were clear (69% vs. 64%) and performed better when grasped objects occluded the wrist cameras (48% vs. 22%).
WhyProposes hierarchical active-gaze policies enabling bimanual manipulation with only stereo cameras, validated across over 1000 real and 1800 simulated trials against strong baselines.
In briefGraph decomposition for decentralized optimization using network-designed message passing
SummaryThe paper studies decentralized minimization of sums of smooth strongly convex functions, where agents communicate only with graph neighbors. It proposes a graph decomposition framework that uses the network to design the optimization method, introducing GATE and its lower-cost surrogate GATE-S. The authors prove linear convergence and show faster convergence on tree-dominant graphs, with rates scaling with tree diameter. Experiments support the predicted scaling and savings in iterations and communication.
WhyProposes a new graph-decomposition framework with topology-aware preconditioning and tree-diameter-dependent linear convergence; experiments support predicted scaling.
In briefLESSER: Choosing Post-Training Data Using Output-Layer Gradients
SummaryThis paper studies how to select data for post-training large language models using gradient-based ranking. Full-parameter gradients require costly backward passes on every sample, so the authors propose LESSER, which approximates them with cheaper output-layer gradients that need only a forward pass. On SFT and RL benchmarks it cuts feature-extraction FLOP cost by 9.7x and 3.0x while tracking full-gradient downstream performance. The authors also find that batches selected this way have aligned gradients, even when individual sample rankings differ.
In briefHow register and high-norm patch tokens in vision transformers play different roles
SummaryThis paper studies the internal tokens of self-supervised Vision Transformers such as DINOv2, focusing on register tokens and high-norm outlier patch tokens that appear in background regions. The authors train sparse autoencoders on both token types and find that register-token features are more tied to high-level semantic concepts, while outlier-token features more often reflect structural, background, and texture patterns. Causal ablations show that disrupting register-derived features reduces representation cosine similarity by 48.17%, compared with only 0.31% for outlier-derived features, supporting functional specialization between the two token types.
In briefChess Language Model That Plays and Explains Its Moves
SummaryThe paper introduces Queen, a 4B-parameter chess-language model that plays at roughly Grandmaster level while explaining its moves and plans. It combines a pretrained chess encoder with an instruction-tuned language model through cross-attention and improves explanations with an iterative distillation method. According to the authors, Elo rose from 1782 to 2697 over seven iterations, surpassing listed frontier models on play strength and puzzle accuracy.
In briefAI combining gene expression and biopsy images to predict breast cancer treatment response
SummaryThis paper proposes a two-stage AI model that first infers gene expression from breast cancer histopathology images, then predicts pathological complete response to neoadjuvant therapy using inferred expression and clinical variables. Evaluated across nine cohorts of 1,412 patients, it reaches a pooled AUROC of 0.79. The authors suggest this biologically informed approach may generalize to data-sparse precision oncology applications.
Jungkyu Park, Dhruva Biswas, Joseph Cappadona et al.
In briefFlowHMR: Capturing Physically Plausible Human Motion from Video
SummaryFlowHMR is a framework that recovers physically plausible global 3D human motion from monocular video. It treats video motion capture as video-conditioned motion generation, pretraining a flow matching model and then post-training it with GRPO using fidelity and physics-tracking rewards. The authors also introduce Wild-4K, a dataset of about 4K internet videos, where the method reportedly achieves an 82.47% physical tracking success rate versus 62.82% for the strongest baseline, GVHMR.
In briefTesting SigLIP2 image model for aerial fire risk sorting
SummaryThis paper tests how well a pretrained SigLIP2 image encoder can classify aerial images into seven fire risk levels. The authors split the public FireRisk training data in a reproducible way and document data sources, preprocessing and model selection. On the validation set, full model adaptation reached 63.05% accuracy and 58.94% macro F1, outperforming a frozen-encoder probe at 55.95% and 50.19%. These are preliminary development results, not yet verified on an independent test set or unseen regions.
In briefTraining population-dynamics models without simulation, using Wasserstein Lagrangian residuals
SummaryThis paper introduces Double-Stitch, a method that learns Wasserstein Lagrangian mechanics without running numerical simulations during training. It penalizes the residual of the equation of motion along a learned population path, which can describe conservative or periodic dynamics that gradient flows cannot. Tested on synthetic, single-cell and ocean vortex datasets, it matches or outperforms gradient-flow methods and simulation-based WLM on most tasks while training 4 to 14 times faster than WLM.
WhyLearns Wasserstein Lagrangian dynamics without simulation via residual penalties, tested on synthetic, single-cell, and ocean data with 4-14x faster training.
Fedor Sergeev, Markus Heinonen, Daniel Waxman et al.
In briefFrugalEvo: Cost-Aware LLM-Guided Program Evolution
SummaryThis paper proposes FrugalEvo, an evolutionary framework where a stronger, costlier LLM explores solution strategies and a cheaper LLM implements and refines the code. It also uses cache-efficient prompts to improve reuse across evolution steps. The authors introduce BA-AUC, the area under the best-so-far score curve over cumulative LLM cost. Across 10 mathematical and systems optimization tasks, FrugalEvo matches or surpasses baselines in final quality, and reports new state-of-the-art results on circle packing at low cost.
In briefRetrieval-centric deep learning in growing nonparametric neural networks
SummaryThis paper proposes a general-purpose deep learning layer that stores a key-value pair for each training data point and retrieves them with attention at inference, so the network grows with the data. The authors show that naively applying linear-attention learning rules to more powerful kernel attention does not yield principled optimization, and they derive functional gradient learning rules for RBF and softmax-like kernels, establishing the retrieval-centric deep learning (RCDL) paradigm. Experiments on image classification and synthetic teacher-student tasks show promising performance and learning efficiency, and replacing linear attention with MesaNet/DeltaNet yields a formal link to optimizers for fixed-size networks.
Maximilian Schlegel, Rajai Nasser, Seijin Kobayashi et al.
In briefPlanning to Learn: A classification loss between cross-entropy and exact policy gradient
SummaryThe paper asks why exact policy gradient underperforms cross-entropy in classification, arguing that it values each update only by its immediate gain and ignores how updates set the starting point for later learning. The authors propose a "horizon loss" that truncates the accumulated error at the learning remaining, moving from cross-entropy toward exact policy gradient as training runs out. According to the paper, on MNIST and on ImageNet with ResNet-50, ResNet-101 and ViT-S/16, it improves top-1 accuracy over cross-entropy at a flat learning rate, with gains growing as label noise increases.
WhyReframes cross-entropy as patient policy gradient and proposes a horizon loss, with a provable allocation model and ImageNet and MNIST experiments.
In briefPivot-SD: Efficient Self-Distillation for Masked Diffusion Language Models
SummaryMasked diffusion language models generate text in parallel, but a few key decisions during denoising largely determine the response, making credit assignment difficult. This paper proposes Pivot-SD, an offline self-distillation framework that supervises only these high-impact decisions (pivots), selected by an information-gain metric. Pivots from successful trajectories are trained with cross-entropy and those from failed ones with targeted unlikelihood. Using only 200 questions with four rollouts each, it improves LLaDA-8B-Instruct over full-sequence SFT and budget-matched diffusion RL baselines on math and code benchmarks.
In briefProAR: Teaching Autoregressive Video Models to Reason Ahead
SummaryAutoregressive video models generate content causally but predict only the next chunk, limiting them to short-sighted generation. This paper proposes ProAR, which adds goal-frame prediction via an asymmetric attention mask and uses future representation self-alignment to guide short-range transitions toward a target outcome. According to the paper, it consistently improves results on diverse visual reasoning benchmarks and surpasses fully trained standard baselines using only 25% of the training steps.
In briefForecasting new policy outcomes from simulated counterfactual rollouts
SummaryThis paper addresses the cold-start problem of forecasting when a new decision policy has just been deployed and no real data yet reflects its effects. The authors roll out the target policy in a simulator across counterfactual scenarios and train a forecaster on the resulting trajectories. In two real inventory-control deployments, the simulator-trained model achieved lower point-forecast MAPE than a model trained on historical real data, and lightweight calibration with early real observations reduced error further.
Angel Wang, Dominique Perrault-Joncas, Alvaro Maggiar et al.
In briefRe-examining input time-frequency representations for multi-pitch estimation in vocal ensembles
SummaryThis paper studies multi-pitch estimation for vocal ensembles, where singers' overlapping pitch ranges and close fundamental frequencies make their harmonics hard to separate. The authors compare the commonly used harmonic constant-Q transform (HCQT) with a linear short-time Fourier transform (STFT) whose frequency bins are fed directly to the model. The linear STFT performs better than HCQT while greatly reducing feature-extraction cost. Longer analysis windows or wider spectral coverage gave no extra benefit, and shorter windows were more effective for time-varying vocal pitches.
In briefMRVQ: One Resident Index for Dimension- and Rate-Elastic Vector Search
SummaryDense retrieval services must switch between embedding dimensions and index bit rates, but keeping separate indices for each setting costs a lot of memory. MRVQ is a post-hoc residual quantizer for frozen embeddings whose maximum-rate code can be truncated by dropping residual stages or embedding coordinates, so one resident artifact serves every evaluated dimension and rate pair. It uses the least RAM among the designs evaluated, but per-rate QINCo2 achieves better quality on FiQA, so MRVQ is not a universal quality winner.
In briefOnboard Detection of Ocean Pollution and Blooms via Satellite Anomaly Spotting
SummaryThis paper presents a marine event detection pipeline for Earth observation satellites with multi- or hyperspectral sensors, targeting hazards such as oil spills, algal blooms, and sediment floods. A self-supervised neural encoder compresses satellite images into a reduced latent space, and an anomaly detection model flags deviations from normal sea patterns. The authors compare it with Isolation Forest, One-Class SVM, and Local Outlier Factor. The pipeline is designed for resource-limited onboard hardware and has been integrated into the ESA Phisat-2 and Microsoft/Thales Alenia Space IMAGIN-e missions.
Thomas Goudemant, Clotilde Szywala, Benjamin Francesconi et al.
In briefVariational inference method improving uncertainty estimates for learned data manifolds
SummaryThe paper studies how to approximate the lower-dimensional manifold where data lies while capturing its epistemic uncertainty. It applies Amortized Structured Stochastic Variational Inference to a Gaussian Process Latent Variable Model, so the latent variable posterior depends conditionally on the inducing point values instead of using a mean-field approximation. The authors report that this more flexible posterior improves several metrics for reconstructing points on the data manifold.
Maksym Tretiakov, Sarah Filippi, Vincent Fortuin et al.
In briefWhen Can a Bandit Stop Relying on Its Past? E-Process-Authorized Thompson Sampling for Non-stationary Environments
SummaryThe paper proposes e-ATS, a multi-armed bandit method that keeps both full-history and discounted Beta states for each arm. An anytime-valid e-process first decides whether the discounted state may be used, and a reversible relevance score then controls its influence. Under a stationary Beta-Bernoulli model, the chance of ever departing from optimistic Thompson sampling is at most a chosen level α_E. Removing this authorization raised normalized dynamic pseudo-regret by 38.4% on a registered suite but lowered it by 7.5% on a literature-derived replay suite, so the evidence governs when adaptation starts rather than guaranteeing it helps.
In briefConvergence of success conditioning for policy optimization
SummarySuccess conditioning is a method for improving decision-making policies in stochastic environments by increasing the probability of actions that lead to successful outcomes. The paper shows that it converges to an optimal policy on a broad class of Markov decision processes. For discounted MDPs, it proves convergence to an ε-optimal policy within O(1/ε^p) iterations, where p depends on problem data. For single-period MDPs, such a policy is obtained within O(log(1/ε)) iterations.
In briefIDRF: Reward fine-tuning for masked discrete diffusion models via inverse distillation
SummaryThis paper introduces IDRF, a framework for reward fine-tuning of few-step masked discrete diffusion generators. It replaces the intractable sequence-level KL penalty with an inverse-distillation regularizer and optimizes the student's own few-step sampling trajectories. The authors report high reward across DNA, image, and text generation with up to 32 times fewer denoising steps than the reference, while mitigating reward hacking and preserving sample quality.
Vladislav Gromadskii, David Li, Samson Gourevitch et al.
In briefBroken scale symmetries in undercomplete linear autoencoders
SummaryThis paper studies how the loss landscape of undercomplete linear autoencoders shapes SGD dynamics. On the PCA solution manifold, finite-stepsize SGD selects a preferred direction of scale drift that favors large decoder weights. The drift is slow and admits an analytically tractable effective description, but it is limited by a finite-stepsize stability boundary.
In briefDo Large Language Models Know Colombian Law? A Reliability Benchmark
SummaryThis paper introduces an expert-validated benchmark of 1,042 items across ten legal areas to test how reliably large language models answer questions about Colombian law. The authors evaluated 15 models on closed multiple-choice and free-text questions. While some models score high on multiple-choice questions, no model's free-text factual correctness exceeded 0.45. The authors conclude that current models need expert supervision for Colombian legal tasks, and that grounding answers in authoritative sources may improve reliability.
Rub\'en Manrique, Michelle Castellanos, Jorge Morales et al.
In briefLoGo: Local-Global Rewards for Long, Consistent Camera-Controlled Video
SummaryThis paper addresses 3D inconsistency in camera-controlled video models, where objects lose permanence and scene structures shift over long generations. It proposes LoGo, a post-training method that blends global rewards with spatially localized rewards to improve credit assignment. The authors report that LoGo improves 3D consistency while preserving camera following and video quality, with gains shown across three base models on DL3DV and the new TrajectoryBench benchmark.
In briefAssigning Credit Where It Matters: Dependency-Aware Policy Optimization for Terminal Agents
SummaryThis paper proposes Dependency-Aware Group Policy Optimization (DepGPO), a reinforcement learning method for terminal agents performing multi-step tasks such as coding and debugging. It builds a command dependency graph from execution traces and traces backward from the resources checked by the task verifier to credit relevant writes and their supporting reads. The method redistributes trajectory advantages across steps accordingly. The authors report that comparative and ablation experiments show improved task performance and training stability on complex terminal tasks.
In briefWorld Embedding Benchmark: A test of whether video representations encode physical information
SummaryThis paper introduces the World Embedding Benchmark, with 8,000 controlled simulation cases across 80 families covering fluid mechanics, solid mechanics, dynamics, and optics and electromagnetism. Each video is paired with simulation-derived physical annotations and evaluated on text-video retrieval, physical-property regression, and multiple-choice pair classification. Pre-trained models show weak retrieval and near-chance pair classification, while lightweight probes recover useful physical information from frozen embeddings. Continual contrastive training on physics-specific pairs improves retrieval and pair classification but degrades regression, and retrieved reference videos improve the physical fidelity of videos generated by MiniMax-H3.
In briefNeutronGym: A Physics-Graded Environment for Designing Neutron Instruments with LLM Agents
SummaryThis paper introduces NeutronGym, an executable environment where language-model agents design neutron scientific instruments using validating tools and McStas ray-tracing. Performance is graded at four levels (syntax, runtime, structure, science) without an LLM judge. Seven models reproduce at most 7 of 16 benchmark tasks. Reinforcement learning on the environment's reward raises Qwen3-8B's success on held-out instances of one task family from 11% to 77%.
WhyIntroduces the first executable neutron instrument design environment with physics-graded rewards; RL gains are shown with ablations, though impact stays within a subfield.
In briefDepth as Time in One-Step Generative Models: Multi-step diffusion denoising reorganized across network depth
SummaryThis paper examines what happens to multi-step diffusion denoising when generation is compressed into a single forward pass. The authors report an observation called "depth as time": the denoising computation spread over sampling steps appears across the layers of one forward pass and can be recovered by decoding intermediate layers with the model's output head. They find this depthwise computation depends on the transport task a flow map is trained for, with MeanFlow showing both denoising and renoising, and state that a MeanFlow SiT-L/2 model can be compressed 16.6× in parameters.
Arnold Caleb Asiimwe, William Yang, Sanghyuk Chun et al.
In briefFALCON: A framework for generating synthetic NL-to-SQL training data that stays model- and database-agnostic.
SummaryThis paper presents FALCON, a framework that generates ambiguity-aware natural-language-to-SQL training data using compact open models. It combines reserved-word SQL seeding and persona-based prompting to create structurally complex queries, and uses alignment-based filtering to keep difficult but valid examples. The authors report that the generated data exceeds existing benchmarks in SQL complexity and language richness, and that it can be produced locally without external APIs.
In briefComputing normal-form correlated equilibria in Markov games
SummaryThis paper gives the first efficient algorithm for computing normal-form correlated equilibria (NFCEs) in finite-horizon Markov games with a fixed number of players, running in time S(AH/ε)^O(n) to find an ε-NFCE. The method applies backward induction over auxiliary stage games, computing a constant-expectation correlated equilibrium at each step. The authors also show PPAD-completeness in some regimes and an equivalence between constant-expectation CEs and NFCEs.
WhyGives the first polynomial-time algorithm for normal-form correlated equilibria in finite-horizon Markov games, plus PPAD-completeness results.
Ioannis Anagnostides, Constantinos Daskalakis, Gabriele Farina et al.
In briefUniIntervene++: An Adaptive Agent That Decides When Humans or Code Should Assist Robot Learning
SummaryThis paper presents UniIntervene++, an adaptive intervention agent that allocates control between a robot's autonomous reinforcement learning policy and assisted behaviors during online learning. It treats the RL policy, trajectory correction, and a code-based policy as options in a unified semi-Markov decision process, and periodically probes the policy through unassisted execution. Across five real-world manipulation tasks, it reports an average success rate of 89.67% and human intervention reduced to 0.77%.
In briefVideo-time priors for physiological emotion prediction on familiar videos
SummaryThis paper asks whether viewer-specific EEG and fNIRS signals improve emotion rating prediction beyond what video identity and playback time alone provide. The authors use a smoothed per-video, per-second median of training labels as a stimulus-only prior and fuse it with a graph-based physiological model. Fusion barely outperforms the prior alone in both cross-validation and an external cohort, so the physiological gain is small. The model supports only offline prediction, and generalization to unseen videos is untested.
In briefDEPICT: A Training-Free Way to Score Whether Generated Images Match Their Text
SummaryThis paper introduces DEPICT, a training-free metric for judging how well a text-to-image output matches its caption. Instead of fixed reference answers, it uses the expected agreement between answers from the image and answers from the caption alone, weighting each question by how decisively the caption determines it. It merges this score with a holistic score to recover context lost during decomposition. Evaluated on five benchmarks and eleven backbones, DEPICT outperforms all training-free metrics.
Vasco Ramos, Sandra Godinho Silva, Joao Magalhaes et al.
In briefBAT-NO: A Transformer Model That Predicts Car Part Crash Performance Under Changing Boundary Conditions
SummaryThis paper proposes BAT-NO, a deep learning surrogate that predicts transient displacement fields and crashworthiness responses of vehicle components, replacing costly finite-element simulations. It evaluates generalisation on a B-pillar framework with variations in geometry, impact position, velocity and support stiffness. BAT-NO achieves the lowest mean displacement error among the baselines, reducing error by 32.6% against the second-best model in the most challenging case. The authors conclude that explicit local and global boundary-condition representations improve crashworthiness prediction across expanded component design spaces.
In briefMastering Atari 2600 Games with Automatically Discovered Options
SummaryThis paper presents Wayfarer, a general, domain-agnostic online deep reinforcement learning agent that discovers options from high-dimensional observations using Laplacian representation learning. According to the authors, these options improve exploration, accelerate credit assignment, and generalise to unseen settings. Wayfarer achieves state-of-the-art performance among single-stream agents on the most challenging Atari 2600 games, with the largest gains in games requiring long-horizon exploration, such as Montezuma's Revenge and Private Eye.
WhyDiscovers options via Laplacian representation learning from pixels, reaching top single-stream Atari results on hard exploration games.
In briefGAMEGO: Training Game-Dev Agents on Synthetic Trajectories Grounded in Real Game Assets
SummaryThis paper introduces GameGo, a framework that turns brief game ideas into detailed Product Requirements Documents based on industry game-development practices, reducing gaps in mechanics and gameplay flow when coding agents build games from sparse queries. The authors release GameGoData, with 55,060 development trajectories across 2D, 2.5D, and 3D games, and GameGoBench, a benchmark of 124 game queries. Their model GameGoCoder, trained on this data, outperforms matched baselines and performs comparably to frontier models on game-development benchmarks.
In briefManifoldSplat: Language-Guided Semantic Shape Editing of 3D Gaussian Head Avatars
SummaryManifoldSplat is an end-to-end framework that edits the shape of animatable 3D Gaussian Splatting head avatars, reconstructed from monocular videos, using natural language instructions. Edits are made within the structured FLAME manifold rather than by directly optimizing the Gaussian cloud, which preserves identity and animation. The method uses DeltaRegion, a per-region disentangled conditional variational autoencoder that produces shape changes in a feedforward manner, followed by a refinement stage for view-consistent details. The authors report state-of-the-art results in localized prompt alignment, geometric coherence, and identity preservation.
In briefPath-integral surrogate for reconstructing private images from multi-step federated updates
SummaryIn federated learning, clients share only model updates, which should reveal less about their data than raw examples. This paper proposes PI-SME, which treats the accumulated multi-step update as a path integral of the gradient field, approximated by Gauss-Legendre quadrature over a learnable Bézier path. Gradient inversion attacks use this to reconstruct private input images from the update. On CIFAR-100 and FEMNIST, the authors report more faithful reconstructions than the strongest surrogate baseline on several inversion metrics and the matching loss.
In briefHazardWeaver: Choosing Valid Scientific Methods for AI Hazard-Analysis Agents
SummaryThis paper introduces HazardWeaver, a framework that helps AI agents for natural hazard analysis select scientific methods that are both applicable and executable with available data and tools. It revises these choices as new evidence and execution results change the analysis state. The authors also release the Hazard Weaver Benchmark, with 141 instances across seven single-hazard and four multi-hazard categories. Experiments report that HazardWeaver outperforms existing agent systems, with the largest gains on tasks that have multiple eligible routes.
In briefChoosing which tensor to cache in few-step diffusion transformers
SummaryThis paper studies how to speed up Diffusion Transformers by caching and reusing a tensor from an earlier step, which becomes riskier after step distillation. It introduces AutoTarget, which runs a few uncached samples to measure the error of each candidate tensor and picks the one with the lowest error. Experiments on distilled image and video DiTs show the best cache target depends on the model, resolution, and solver. The method reduces DiT evaluations and cache storage while keeping quality close to uncached generation.
In briefHyperBrowseComp: A Hard Multilingual, Multimodal Benchmark for Web-Browsing Agents
SummaryThis paper introduces HyperBrowseComp, a benchmark of 423 manually written and human-validated questions in 13 languages. The questions require finding obscure evidence, following multi-step clues, and inspecting videos, scanned documents, images, or maps on the open web. Easier questions were filtered out by testing models without internet access. Several models were evaluated under a common agent protocol, with a human evaluation on a sample of questions.
Alham Fikri Aji, Faiz Rizki Ramadhan, Zayd M. K. Zuhri et al.
In briefLearning to judge whether mmWave radar can read a heartbeat
SummaryThis paper studies how to decide, for each millimeter-wave radar measurement, whether it contains a readable heartbeat signal for reliable contactless heart-rate estimation. The authors build a controllable multi-scatterer FMCW simulator to generate training data with automatic observability labels. They propose HEAR, a compact dual-task Transformer that jointly predicts an observability score and heart rate. Trained only on simulated data, HEAR transfers zero-shot to two public 60 and 120 GHz datasets, where the score enables selective prediction.
In briefWriterslogic's CLEF 2026 SimpleText entry: multi-candidate LLM simplification and complexity spotting
SummaryThis paper describes the Writerslogic team's submission to the CLEF 2026 SimpleText shared task, covering text simplification and complexity spotting. For simplification, GPT-4o-mini generates five candidates per sentence, and the best is chosen with a reference-free scoring heuristic. For spotting, hallucination detection is framed as natural language inference using a fine-tuned DeBERTa-v3-large model. The authors report top-ranked results on both tasks using biomedical Cochrane review text.
In briefWriterslogic's PAN 2026 systems: process-based features for robust detection under domain shift
SummaryThis paper describes Writerslogic's systems for three PAN at CLEF 2026 shared tasks, arguing that feature robustness under domain shift depends on how much training and test distributions overlap. On Reasoning Trajectory Detection, the team ranked 1st in source detection (0.85 macro F1) and 3rd in safety classification (0.66 macro F1). Their best Voight-Kampff Generative AI Detection configuration scored 0.891 on the PAN 2026 test set. The Multi-Author Writing Style Analysis run did not reach official evaluation due to a platform error, so only the design and a priori predictions are reported.
In briefKnowledge or Calculator? Splitting the Skill Premium in Verifiable Financial AI Agent Workflows
SummaryThe paper introduces FinSkillBench, a suite of 2,603 point-in-time episodes across 12 financial subtasks, scored by deterministic verifiers. Curated skill packages raised mean scores by about 16.2 points, while skills generated within a single episode added only 0.5 points. Separately, executable tools contributed more than procedural documents, and their combination was subadditive, with effects varying by workflow type.
Jermyn Zhen Yong Bek, Zhuang Qiang Bok, Zhongtian Sun
In briefCephalonauts One: a deep fMRI dataset for decoding natural speech in the brain
SummaryThis dataset records whole-brain 3T fMRI activity from three healthy subjects listening to podcasts, with about 30 hours of data per subject, making it the deepest such dataset using naturalistic speech. It pairs brain recordings with the podcast audio, transcript annotations, and stimulus embeddings. The authors also define a decoding benchmark in which a model must match fMRI activity from a held-out session to the correct time-aligned audio segment, and they provide splits, metrics, and baseline decoders. A scaling analysis shows decoding performance keeps improving as training data per subject increases.
Antoine Collas, Louis Jalouzot, G\'eraud Ilinca et al.
In briefGet a GRIP, this will be a long TRIP: A quantifiable framework for testing over-squashing in long-range graph learning
SummaryThe paper argues that claims about over-squashing and long-range interactions in GNNs are only trustworthy if the benchmarks truly require long-range information. It introduces four verifiable axioms that such tasks must satisfy, and proposes TRIP and GRIP, procedures that turn any graph into a provably long-range task. The authors audit four common long-range benchmarks against these axioms and find that a popular curvature measure is uncorrelated with GNN performance.
WhyProposes four verifiable axioms and a constructive framework with a per-range error lower bound, backed by formal proofs and audits of common long-range benchmarks, substantively advancing GNN long-ra
Ferran Hernandez Caralt, Simon Heilig, Adri\'an Bazaga et al.
In briefObjects Without Morphisms: What Math LLMs Fail to Represent
SummaryThis paper studies how large language models translate mathematical statements between the terminology of neighbouring subfields when no external checker exists. The authors introduce a judge-free instrument that measures whether a rewrite states the hypothesis implicit in its source. Across seven models, translations toward general framing widened the quantifier domain in 60.6% of cases and never narrowed it, while concrete framing usually narrowed it, and the most capable models widened least.
WhyIntroduces a new diagnostic of generality-level fidelity in math translation, supported by blind coding, planted controls, and pre-registered held-out replication across seven models.
In briefSelf-Improving Synthesis Harness for Harder Reasoning Data
SummaryThe paper proposes co-evolving tasks and the synthesis harness, so the procedure that generates reasoning problems improves itself rather than only reusing generated problems. Across mathematics, coding, and science, mean solver accuracy drops from 100.0% to 54.8% over fourteen evolution rounds. A 27B student fine-tuned on 10K synthesized mathematics examples reaches 62.5% mean-16 accuracy on APEX.
In briefZeroMAG: Zero-Shot Multimodal Adapter Generation for Plug-and-Play EEG Models
SummaryThis paper presents ZeroMAG, a framework that extends a frozen EEG foundation model to multimodal recordings that include companion physiological signals. It generates adapter weights from unlabeled target recordings without target labels or target-side optimization. Across six held-out target datasets and three backbones, it improves balanced accuracy by 7.22 points over EEG-only inference and comes within 0.50 points of supervised multimodal adaptation on average.
In briefTheory of deterministic probability-flow ODE samplers for diffusion models
SummaryThe paper examines when probability-flow ODEs, used as deterministic samplers for score-based diffusion models, define a well-posed flow. It shows that rewriting the Fokker-Planck equation does not by itself guarantee this, and gives sufficient conditions for a regular flow and two-sided divergence control for reverse sampling and invertibility. It also analyzes stability for learned scores and gives a sphere example where the deterministic flow becomes singular as noise vanishes.
In briefDuoMatching: A Method That Improves Fast Video Generation by Matching Both Whole-Video and Single-Frame Distributions
SummaryThis paper proposes DuoMatching, a distribution matching framework for distilling video generators that adds a frame-level marginal matching objective on top of existing joint matching. Frame-level supervision is provided by an image generator, with LatentBridge resolving the latent mismatch between the video student and image teacher, and Latent Variation Sampling spreading that supervision across temporal segments. According to the experiments, the method improves visual quality, composition, and semantic alignment while largely preserving motion dynamics. Human evaluators preferred DuoMatching over all evaluated baselines at rates above 80%.
In briefAbdominal ultrasound simulation from semantic labels via paired label-to-physics image translation
SummaryThe paper proposes a two-stage learning pipeline that generates realistic abdominal ultrasound images from anatomical segmentation labels without needing patient-specific CT volumes at inference time. Stage I compares Pix2Pix and the Semantic Diffusion Model (SDM) in predicting simplified CT-based ray-casting ultrasound images, with SDM performing better on morphological metrics. Stage II refines these outputs with segmentation-guided CycleGAN (SG-CycleGAN). CT-derived segmentations and ray-casting simulations are still required to train Stage I.
Santiago Vitale, Duilio Deangeli, Ignacio Larrabide et al.
In briefAutonomous robot navigation for reaching brain-computer interfaces through blood vessels
SummaryThis paper demonstrates robotic navigation in cerebral veins in vitro to deliver endovascular brain-computer interface devices without craniotomy. Soft Actor-Critic controllers were trained in simulation for two tasks from the right internal jugular vein to the superior sagittal sinus. Simulated success reached 85.6% and 98.4% in the training anatomy and 42.0% and 91.6% in an unseen anatomy, and 14 of 20 physical runs succeeded.
Harry Robertshaw, Weijie Qi, Nikola Fischer et al.
In briefComparing author representation methods for zero-shot authorship attribution
SummaryThis paper studies how representing authors affects zero-shot authorship attribution, where no task-specific training data is available. It compares a label-only prompting baseline with three author representations: representative writing samples, LLM-generated style descriptions, and style embeddings (LISA), and proposes a two-stage embedding framework. Label-only attribution performs poorly, while adding author-specific representations consistently improves results. The proposed two-stage LISA framework performs best overall, while LLM-generated descriptions are more compact but lower in attribution performance.
Nudrat Habib, Tosin Adewumi, Sana Sabah Al-Azzawi et al.
In briefHow divergence choice controls entropy in LLM distillation: a theoretical and empirical paper
SummaryThis paper studies large language model distillation from an entropic perspective, examining how the student's entropy depends on the data and the divergence defining the objective. The authors prove that forward KL inflates the student's entropy above the teacher's, with cross-entropy training as a special case, and verify this quantitatively in pretraining and supervised finetuning. Reverse KL instead deflates entropy, while interpolating between the two changes entropy smoothly early in training and abruptly at convergence. The lower entropy of on-policy distillation is attributed to token-level reverse KL rather than on-policy sampling.
WhyProves forward KL raises student entropy above teacher's and shows divergence acts as implicit entropy regularizer, with quantitative pretraining and finetuning verification.
In briefBeyond Trained Models: A Benchmark of Compiled GNNs for Testing GNN Explainers
SummaryExplainers for graph neural networks are often judged by whether their explanations match a planted ground-truth motif, but the paper shows this assumes the model actually uses that motif, which fails on several common benchmarks. The authors introduce Gracr, a compiler that turns graded modal logic formulas into GNN weights, so the model's behaviour and correct explanation are known exactly. Using this, they build GracrBench and test eleven explainers on six tasks. Most explainers prove not robust to indirect influences or to alternative implementations of the same formula.
WhyIntroduces the first compiler from graded modal logic to GNN weights with exact ground truth, validated across eleven explainers and six tasks.
Steve Azzolin, Francesco Paolo Nerini, Stefano Teso et al.
In briefA Method for Composing Verifiable Atomic Insights into Reports Generated from Tables
SummaryThis paper addresses automatically generating analytical reports from relational tables. The authors propose ComInsight, which treats insight discovery as composing atomic evidence: it enumerates valid atomic insights, organizes them into a multi-relational insight graph, and fuses them into higher-order conclusions using composition operators. Each conclusion comes with executable SQL and fine-grained provenance. The method is reported to outperform strong baselines on three benchmarks: InsightBench, DDR-Bench, and T2R-Bench.
In briefFrom Benchmarks to Production: A Text-to-SQL System for Complex Financial Data
SummaryGeneral Text-to-SQL systems perform well on academic benchmarks but fall below 50% on production financial databases, where concepts are stored as opaque integer keys. The paper presents FLINT, a finance-specialized system with a lookup agent, embedding-based retrieval of query templates from an expert-authored bank, and foreign-key-based schema linking. Evaluated on 359 questions across two production financial schemas, FLINT outperforms several state-of-the-art baselines using the same LLM and is deployed in a financial data retrieval service.
Arijit Sehanobish, Bruno Gomes Coelho, Guillaume Michel et al.
In briefNovel view synthesis across mixed camera types, including fisheye and panoramic
SummaryThis paper studies feed-forward novel view synthesis from sparse posed images taken by mixed camera types such as perspective, fisheye, and panoramic cameras. The authors argue that existing camera encodings miss the projection-dependent ray distribution within each visual token, and they propose local raymaps and projection-aware 2D RoPE to address this. On ScanNet++ with heterogeneous cameras, the method outperforms camera-conditioned baselines under mixed-camera evaluation and generalizes zero-shot to panoramic views.
In briefXGenAct: A Robot World Model That Predicts Depth and Segmentation Alongside Video
SummaryXGenAct is a world action model for robot control that encodes RGB observations, actions, depth, surface normals, and functional segmentation as RGB videos. It uses one video diffusion transformer and one training objective to predict these signals over time, without task-specific heads. On held-out RLBench tasks it improves closed-loop success over RGB-only training, reaching 52% in a five-task external comparison versus 26% for the strongest baselines. It also predicts future depth and segmentation more accurately than pipelines that generate RGB first and then apply a frozen perception model.
In briefLearning from Repaired Reasoning: Root-Cause-Guided On-Policy Distillation
SummaryThis paper proposes Root-Cause-Guided On-Policy Distillation (RC-OPD), a method for improving reasoning training in language models. It addresses the mismatch between reference-based guidance and the student model's actual reasoning errors by locating the earliest substantive error, developing a local correction, and anchoring valid prefixes to the repaired intermediate results. Experiments across multiple datasets and model scales show that RC-OPD mitigates reasoning mismatch and the distillation trap, yielding substantial performance gains.
In briefProgressNet: Step-by-step sketching and prompting with a frozen text-to-image model
SummaryProgressNet is a training-free framework that lets a frozen text-to-image model follow a drawing session as it unfolds, handling stroke additions, erasures, and prompt revisions at about one second per turn. It uses no new parameters, relying on three inference-time mechanisms that exploit existing model features. On FS-COCO, the FID of the FLUX+ControlNet baseline doubles as the sketch goes from 10% to 100% completion, while ProgressNet's barely changes. Users preferred it over five competitors, most widely for erasure.
In briefWeave Forcing: Reusing Past Video Memories to Keep Characters and Backgrounds Consistent
SummaryThis paper presents Weave Forcing, a training-free framework for reusing earlier shots when generating interactive long videos. An LLM splits each prompt into character and background parts and picks suitable historical references for each, while masked memory weaving selects relevant content from compressed memories. Coverage adaptive RoPE adjusts temporal offsets and memory retention based on how much of the reference is available. Experiments report improved subject and background consistency across shots with competitive visual quality and text alignment.
In briefEfficient reasoning training does not always reduce CoT faithfulness or monitorability
SummaryThis paper examines whether training language models to reason with fewer tokens makes their chain-of-thought (CoT) less faithful to their actual decisions. The authors fine-tune several models using three methods that apply length pressure differently and measure faithfulness and monitorability. Faithfulness drops in most settings, mainly because the trained models become less consistent. Monitorability holds up better, as models still acknowledge input influences on their answers even when the CoT is much shorter.
Samuel Lewis-Lim, Xingwei Tan, Mario Sanger et al.
In briefAn automated, reproducible workflow for finding cracks and assessing damage in fusion materials
SummaryThe paper presents a workflow built in the Galaxy scientific workflow environment that automatically detects cracks in scanning electron microscope images and quantifies damage, while keeping intermediate results for reproducibility. It is designed to work across different tungsten grades, microstructures, magnifications and damage states without per-image parameter tuning. The workflow is demonstrated on an electron-beam thermal-shock dataset of 418 images from 114 experiments and defines a crack-density descriptor.
Rinkle Juneja, Viktor Reshniak, Richard K. Archibald et al.
In briefTuning guidance weights automatically for diffusion and flow-matching posterior sampling
SummaryThis paper addresses training-free posterior sampling (Plug-and-Play) methods that use pretrained diffusion or flow-matching models to solve inverse problems. Such methods depend on guidance weights that are usually set by heuristics. The authors show that the conditional training objective at each time step is a least-squares problem, so the weights can be optimized offline as a two-dimensional linear least-squares problem. The approach reportedly improves reconstruction and lets diffusion samplers cut steps from 1000 to 50 without significant quality loss..
Liam Moroy, Jean-Fran\c{c}ois Giovannelli, Yoann Altmann et al.
In briefKVE-KD: A knowledge distillation method for vision-language models that focuses on key visual evidence
SummaryThis paper proposes KVE-KD, a knowledge distillation framework for compressing vision-language models so they can run on resource-constrained devices. It uses the teacher model to identify task-relevant visual tokens and focuses feature distillation on them while suppressing background information. The authors report that it outperforms existing cross-modal distillation methods on six benchmarks, with the largest gains on complex reasoning and fine-grained visual understanding, and no added inference-time overhead.
In briefDo deep models beat simple baselines for tabular data generation at small training sizes?
SummaryThis preregistered study benchmarks seven tabular data generators, including CTGAN, TVAE, and TabDDPM, against simple baselines across eight public datasets subsampled from 200 to 20,000 rows and four small clinical datasets. Performance is measured by the AUROC of classifiers trained on synthetic data and tested on real data, across 2,220 runs. In 23 of 24 dataset-model pairs, no deep model beat the best trivial baseline beyond seed noise at any size tested. The authors also report that the predicted instability of model rankings at small sizes was not observed.
In briefDetect and Suppress: A Defense Against Adversarial Patches for Robot Models
SummaryThis paper studies how adversarial patches disrupt Vision-Language-Action (VLA) robot models by altering visual inputs. Using a sparse autoencoder, the authors identify an internal feature strongly correlated with patch presence. They suppress this feature at inference only when a linear probe detects an attack, avoiding fine-tuning. On LIBERO-10, conditional intervention improves success under intermittent attacks, while continuous intervention degrades performance.
Yukiya Horiba, Koshiro Aoki, Shunsuke Yasuki et al.
In briefMARO: Recent-First Recurrent Ordering for Time Series Forecasting
SummaryThis paper proposes MARO, a model for time series forecasting that processes the look-back window from the most recent patch to the oldest. The latest patch anchors the latent state and conditions each recurrent step. A single shared module is reused at every step, and intermediate states let the forecast head weigh short and long history separately. The authors report state-of-the-art results on long-term and short-term forecasting across multiple real-world datasets.
Jung Min Choi, Ngoc Son Le, Ibram Abdelmalak et al.
In briefDuoTS: Dual-Context Analog Retrieval for Long-Term Time Series Forecasting
SummaryThis paper proposes DuoTS, a time-series forecasting model that first makes a base prediction and then refines it one future patch at a time. Each refinement blends a current-context view of recent dynamics with a detail-context view of retrieved historical analogs and their subsequent trajectories. The authors report state-of-the-art results on multiple real-world datasets, with ablations supporting each context's contribution. The refinement mechanism is described as model-agnostic and can be integrated into existing forecasting models.
Jung Min Choi, Ngoc Son Le, Ibram Abdelmalak et al.
In briefAREX: A training-free sampler that handles part of flow matching dynamics analytically for faster few-step generation
SummaryThe paper introduces AREX, a sampler for pretrained flow matching models that needs no retraining. It uses the mean and covariance of the target distribution to handle the affine part of the sampling dynamics analytically, and integrates only the neural residual term numerically. According to the authors, AREX improves sample fidelity in the few-step sampling regime on image and text-to-image generation tasks.
In briefSingle-pass uncertainty heads for detecting claim-level hallucinations in Persian medical models
SummaryThis paper proposes lightweight uncertainty heads that predict hallucination risk for Persian medical language models from a frozen generator's internal signals. The heads need only a single forward pass, avoiding costly repeated sampling. They were trained on paired claim-level datasets built from Gaokerena-V and Gaokerena-R responses to 168 Iranian medical entrance exam questions. On held-out tests, they reached PR-AUCs of 0.4820 and 0.4652 and ROC-AUCs of 0.7852 and 0.7810, with no retrieval or repeated sampling needed at inference.
Mehrdad Ghassabi, Pedram Rostami, Hamidreza Baradaran Kashani et al.
In briefMetropolis-Hastings Outperforms Importance Resampling for Policy Composition
SummaryThis paper studies decoding-time policy composition, which combines reward-specific language model policies at inference to adjust trade-offs between rewards without retraining. Standard methods combine next-token probabilities and thus introduce sampling bias. The authors analyze an iterative correction based on independence Metropolis-Hastings (MH) and show that, for the same rollout budget, it is at least as close to the target distribution as sampling-importance-resampling (SIR) under any convex f-divergence. They also derive an improvement bound over the uncorrected decoder and experiments in enumerable and LLM-scale settings.
In briefMobiAgent: A Self-Improving Agent Framework for Long-Horizon Mobile Robot Manipulation
SummaryThis paper introduces MobiAgent, a dual-loop agent framework for mobile manipulation tasks that span many steps. The inner loop separates high-level vision-language planning from low-level control using composable atomic skills executed by flow-matching experts sharing one backbone, while the outer loop autonomously segments, verifies, and clusters rollouts to refine the skill library without human annotation. The authors report that MobiAgent outperforms π0.5-TA by 22.5 percentage points on BEHAVIOR-1K, and that autonomous data recycling raises success rates on RoboCasa and Astribot S1.
In briefSingle or Multiple Policies for Phase-Structured Reinforcement Learning?
SummaryThis paper studies non-stationary reinforcement learning problems that decompose into phases, comparing separate policies for each phase against one state-augmented policy shared across phases. The authors show theoretically that the shared policy can match the performance of any multi-policy solution. They propose a regime-based phase decomposition method, based on transient dynamics relative to quasi-stationary durations, to determine which approach performs better. Numerical experiments test four hypotheses about phase duration, phase heterogeneity, data requirements, and transition dynamics.
In briefFed-ADApt: A federated method that adjusts model depth to each site's compute for medical image segmentation
SummaryThis paper proposes Fed-ADApt, a federated learning framework for UNet-based medical image segmentation that adapts network depth. Each hospital can train according to its local compute budget, and the shared model supports dynamic depth selection at deployment. The method was tested on 2D retinal fundus disc segmentation and 3D brain tumor segmentation across multiple sites. The authors report matching full-resource FedAvg in 3D, a 4.7% average Dice reduction in 2D, and lower inference and training costs.
Abhijeet Parida, Zhifan Jiang, Pooneh Roshanitabrizi et al.
In briefRelearning Savings for Specific Facts in Test-Time Continual Learning
SummaryThis paper asks whether a language model that has lost access to a fact still retains a benefit from having learned it. The authors compare three matched training histories on Qwen2.5-1.5B: learning the correct answer, exposure without it, and learning a wrong answer. After continual fine-tuning without rehearsal, the model that learned the correct answer relearned the target faster than the control. The authors conclude that current output accuracy can underestimate the lasting effect of earlier experience.
WhyIntroduces relearning advantage as a measure of lasting binding effects, with matched controls, but evidence rests on one small model and impact is subfield-limited.
In briefUniDynamics: Fusing Event and RGB Data to Generate Future 4D Dynamic Scenes
SummaryUniDynamics is a diffusion-based framework that generates future RGB, depth, and optical flow 4D dynamic scenes from a single event-RGB pair. It does not require long histories or control priors, and it explicitly models future motion fields. The method uses an Event Latent Enhancement module and a Perceptual Dynamics Space inside a multi-scale U-Net to improve geometric and motion consistency. The authors report state-of-the-art results on VKitti2 and DSEC, especially under high-speed motion blur.
In briefAutomated 3D modeling of field maize from point clouds using a vision-language model
SummaryThis paper presents an automated pipeline that builds editable procedural 3D models of field-grown maize from raw point clouds, without manual tuning or species-specific training data. A multimodal vision-language model annotates leaf midlines, and deterministic geometric steps and differentiable NURBS fitting turn these into refined leaf surfaces. On 100 maize plants from the MaizeField3D dataset, the whole-plant median Chamfer distance was 5.4 mm, and 99.4% of curated reference leaves were recovered.
Mozhgan Hadadi, Talukder Z. Jubery, Adarsh Krishnamurthy et al.
In briefKeeping Colon Segmentation Connected: A Three-Stage Method for 3D CT Scans
SummaryThis paper proposes a three-stage segmentation pipeline that preserves topology, aimed at fixing disconnected colon predictions from deep learning methods. The pipeline performs initial deep learning segmentation, then centreline bridging to reconnect separated regions, followed by a reconstruction stage to refine continuity. Evaluations on the TotalSegmentator and RAOS datasets using overlap, distance and topology metrics show improved structural consistency while maintaining segmentation accuracy.
Deshan Kalupahana, Sonit Singh, Praveen Ravindran et al.
In briefWhen Is Accuracy Evidence? A Unified Theory of Generalisation and Validation
SummaryThis paper proposes γ-CUBV, a framework that replaces K-fold cross-validation accuracy with conservative upper bounds on true risk, controlled by a cumulant envelope of the generalisation gap. For equicorrelated folds it derives an effective number of folds, Keff = K/[1+(K-1)ρ], showing that more folds do not always add statistical evidence. Experiments on heterogeneous Gaussian mixtures compare K-fold validation with corrected full-sample resubstitution using coverage and tightness.
In briefSYNLAT: Syntax-Aligned Compression of Reasoning Chains into Text and Latent Representations
SummaryLong chain-of-thought reasoning produces many output tokens, which is costly. This paper introduces SynLat, a framework that places compression boundaries using non-overlapping Syntax-Aligned Units so that phrases, formulas, and local derivations stay intact. A teacher conditioned on the answer builds KEEP/LATENT targets for a single student that, at inference, produces mixed reasoning from only the question and a requested compression level. The authors report that SynLat matches or exceeds the strongest baseline in all 12 task-group aggregates and strictly leads in 11 across two Qwen3 scales and three compression levels.
In briefNear-zero monitor readings do not show that behavior is controlled
SummaryThe paper studies monitors used inside reward-based training to curb reward hacking in code generation. Three monitors that pass the same offline check all yield near-zero readouts, yet the actual share of hacking varies widely by random seed. The authors conclude that low monitor values are not evidence of behavioral control and that an out-of-band behavioral check is needed.
WhyShows that identical low monitor readouts can hide very different reward-hacking shares across seeds, a useful caution for monitor-in-the-loop training.
In briefBeyond Random Splits: Testing Drug-Target Affinity Models Under Realistic Distribution Shifts
SummaryThis paper evaluates drug-target affinity (DTA) models, which help rank candidate compounds before lab screening. The authors built a dataset of 718,800 unique drug-protein pairs from ChEMBL and BindingDB and compared a baseline with 12 architectures. Error rose sharply under protein-level and dual distribution shifts, and model rankings changed depending on the shift used. The authors conclude that DTA benchmarks should match the chemical and target shifts expected at deployment.
Minjae Chung, Clara Li, Malar Paavai Muthukumaran et al.
