In briefAnomaly-ranked confirmation can slow learning of new classes
SummaryAnomaly detectors rank unfamiliar examples for expert confirmation, and the confirmed examples train a classifier to recognise a new class. Across five benchmarks from physics, medicine and computer vision, the examples flagged by the detector are the ones the classifier learns first. In three systems, confirming the highest-ranked examples of a new class yielded lower recognition than confirming the same number of random examples.
In briefR$^2$-Omni: Reducing redundancy in omni-modal AI models for long-form text generation
SummaryThis paper addresses the high computational cost and memory use of omni-modal large language models that process audio-visual inputs. It proposes R$^2$-Omni, a training-free framework with two parts: Structure-Aware Token Consolidation merges redundant token representations using video spatiotemporal similarity and audio local continuity, and Semantic-Sensitivity Guided Precision Allocation assigns value precision to retained tokens. On long-form generation benchmarks, the authors report an average 5.0x memory reduction, 97.98% of baseline generation quality preserved, and a 1.48x end-to-end speedup.
In briefA Common Language for Dynamic Graphs: Enabling Foundation Models via Interaction Tokenization
SummaryThis paper proposes CL-DGFM, a foundation model for temporal link prediction on dynamic graphs. Its Dynamic Interaction Tokenizer encodes interaction memory, historical connectivity, and ordered destination transitions into calibrated tabular inputs, so existing predictors can use dynamic histories without an extra trainable graph encoder. A pretrained tabular foundation predictor is trained across diverse synthetic dynamics and transfers to real graphs without updating target-graph parameters. The authors report competitive results on TGB and TGB-Seq.
In briefHyperFlow: Class-Incremental Semantic Change Detection via Hyperbolic Flow
SummaryThis paper addresses semantic change detection in remote sensing, where a model learns to recognize new types of land-cover change over time instead of using a fixed set of categories. It represents land-cover classes in hyperbolic space to capture their hierarchical relationships and introduces HyperFlow, which models change as an anomaly within a learned semantic flow using Busemann prototypes. The authors also release HiChange, a high-resolution dataset with fine-grained hierarchical labels, and report strong change detection and segmentation results across incremental steps.
Manon B\'echaz, L\'eo Buecher, Emanuele Dalsasso et al.
In briefPredicting RLHF policy responses to reward-model edits before training
SummaryThis paper studies how an RLHF policy changes when its reward model is edited. The authors propose forecasting the policy's response before training, using samples from the current policy and the scores of the old and new reward models. In 54 controlled RL runs, the covariance-based forecast predicts the direction of most feature changes and holds under GRPO and real reward-model upgrades.
WhyThe authors predict policy responses to reward-model edits before training via within-prompt covariance, reporting pre-registered validation across 54 RL runs and real upgrades.
In briefScienceCUA: A framework that builds and evolves skills for scientific software agents
SummaryScienceCUA is a two-stage framework that turns scientific software documentation into reusable executable skills and lookup references for computer-use agents. It improves skills by comparing successful and failed execution trajectories on the same task, then updates the skill library after repeated validation in real environments. The authors also introduce two improved benchmarks, ScienceBoard-Verified and ScienceBoard-VE, for reliable evaluation. Experiments report average task success gains of up to 7.23 points from initial skills and references, and up to 17.12 points after skill self-evolution.
In briefAxion 1.0: A simple, fully open agent foundation model
SummaryThis paper introduces Axion 1.0, a fully open agent foundation model trained in three stages: two lightweight supervised fine-tuning stages followed by reinforcement learning. The first stage builds self-contained reasoning, the second extends it to tool-dependent reasoning, and reinforcement learning strengthens long-horizon reasoning and interaction with external tools. The authors report strong results on a broad set of agentic benchmarks and release the full recipe, including data construction, verification, prompts, and training code.
In briefCalibrating multimodal test-time adaptation under modality reliability and availability shifts
SummaryThis paper studies multimodal test-time adaptation, which adapts pretrained multimodal models to distribution shifts at inference using only unlabeled test data. The authors propose Adaptive Probabilistic Gaussian Calibration (AdaPGC), which extends category-conditional Gaussian modeling to modality-specific and multimodal fusion representations. It uses the resulting probabilistic predictions to calibrate source predictions and guide reliability-based rectification, and recovers unavailable fusion representations from joint distributions. Experiments under diverse modality corruption and missingness conditions show improved robustness under both reliability and availability shifts.
In briefCL3DOD: Balancing Memory and Learning in Outdoor Class-Incremental 3D Object Detection
SummaryThis paper provides the first unified analysis of class-incremental 3D object detection for outdoor driving across multiple benchmarks, KITTI and nuScenes, under two-task and multi-stage settings. It introduces CL3DOD, which fills in missing labels for previously learned classes using the previous-stage detector, preserves old-class responses via query-wise knowledge distillation, and applies EMA-based temporal consistency regularization. Experiments on both benchmarks show CL3DOD generally achieves a better balance between retaining old classes and learning new ones than controlled incremental baselines.
Muhammad Ahmed Ullah Khan, Khalil Ahmed, Didier Stricker et al.
In briefAnchorComplete: Rebuilding Full 3D Scene Meshes from Partial Point Clouds
SummaryThis paper presents AnchorComplete, a framework that reconstructs complete 3D scene meshes from partial point clouds while keeping the observed geometry. It builds on TRELLIS.2 and combines partial-point conditioning with coarse-to-fine, metric-scale generation to fill in unobserved parts. Experiments on Replica and SCRREAM report that it outperforms strong baselines, especially in sparse-view settings.
Weirong Chen, Ismail El-Yamany, Ganlin Zhang et al.
In briefWhat structurally breaks an AI agent? A difficulty audit for stateful tool workflows
SummaryThe paper presents a small executable instrument that varies dependency depth, candidate count, active state width, and access to state snapshots while keeping the number of required operations fixed, to diagnose why agents fail. It finds that increasing dependency depth changes both scheduling freedom and the number of reachable completion states, so these controls are coupled. The authors report results from large generated task sets and a follow-on design, while stating that their small local and descriptive studies do not support general capability claims.
In briefCERBO: A method for training gradient-boosted trees that offers certified recourse at a requested rate
SummaryThis paper presents CERBO, a method for training gradient-boosted decision trees so that people receiving an adverse decision can reverse it with an affordable change. Given a requested recourse rate, the method adjusts its weight after each tree until the certified rate meets the target, and a calibration step extends the guarantee to new individuals. In the authors' experiments, CERBO raised recourse rates at equal accuracy and improved the accuracy-recourse frontier across the tested settings.
WhyThe authors present a GBDT training method targeting a requested recourse rate with certified actions, reporting gains over RABIT on FICO, Bail, and six datasets.
Farhan Shahriyar Hossain, Taufikur Rahman Fuad, Md. Zarif Bin Hasnat et al.
In briefCan coding agents keep improving across tasks without forgetting?
SummaryThis paper studies whether coding agents can continually evolve their harness, the external framework around the model, across multiple tasks without changing the foundation model's parameters. The authors organize three software engineering datasets into six evolution directions and test two representative harness evolution methods. They find that evolving on a later dataset can substantially degrade performance on earlier ones; for example, Meta-Harness's SWE-Bench score fell from 94.0% to 32.0% after further evolution on FeatureBench.
In briefCutting SNAP Payment Errors Cheaply with AI: A Simulation Study
SummaryNew federal legislation will make states with high SNAP payment error rates pay part of benefit costs, which could exceed a billion dollars a year for some states. The study uses simulations based on public SNAP quality control data to compare risk-based preauthorization review with AI that flags specific possible discrepancies for reviewers. Under its main assumptions, risk-based review alone costs about seven times as much as combining AI flags with risk scores, and this cost advantage holds across many plausible assumptions.
In briefCFD-driven symbolic regression for corrected RANS transition models
SummaryThe paper proposes a CFD-driven sparse symbolic regression framework that learns interpretable closed-form corrections to the production and destruction activation functions in the intermittency equation of the k-ω-γ model. Each correction combines a polynomial of a local pressure-gradient parameter with a bounded exponential of local flow features, avoiding ad hoc limiters. Trained on wall-resolved LES data from ERCOFTAC T3 flat-plate cases, the model achieves comparable or slightly better accuracy on training and unseen cases, and without retraining it improves heat-transfer predictions on the transonic LS89 turbine cascade.
In briefDynSL: Adaptive Model Training Across Mixed Edge Devices
SummaryDynSL is a framework that dynamically splits a machine learning model's layers between edge devices and a server. It continuously monitors system metrics and periodically solves an optimization problem covering client and server computation and communication, reconfiguring the split when beneficial. Experiments on a testbed of 24 heterogeneous edge devices show training time reduced by up to about 30% versus the strongest adaptive baseline and by more than 50% versus a suboptimal uniform split. Accuracy was not reduced and sometimes improved.
Boris Radovi\v{c}, Istemi Ekin Akkus, Marco Canini et al.
In briefConvergence of standard Adam optimizer when gradient noise has heavy tails
SummaryThis paper studies how the standard Adam optimizer converges when the noise in stochastic gradients is heavy-tailed, covering both expected risk and empirical risk minimization. Under Hölder smoothness and a bounded p-th moment assumption, the authors give explicit upper bounds on the expected squared gradient norm and show the convergence term decays more slowly as p decreases. They also propose a multistage Adam with practical learning rate schedules and show that larger batch sizes reduce index-sampling noise and improve convergence guarantees. Numerical experiments support these theoretical findings.
In briefCISA: Speeding Up Long-Context Models by Retrofitting Sparse Attention
SummaryThis paper introduces Chunk-Indexed Sparse Attention (CISA), a method that adapts pretrained dense-attention language models without any gradient updates. It selects fixed-size key-value chunks using compact summaries of existing keys and shares those selections across queries in a group to cut long-context inference cost. Tests on NVIDIA Nemotron 3 Nano and Ultra show core attention up to 13.0x faster and end-to-end throughput up to 2.61x that of dense attention. Scores on long-context benchmarks stay close to dense attention.
In briefWritten Once, Read Differently: How Language Models' Own Answers Steer Later Ones
SummaryThe study examines 22 open-weight language models (0.5B to 72B parameters) by fixing each model's answer to a yes/no question and measuring how it shifts the answer to a second question about the same entity, across 3,600 question pairs about 120 entities. Each answer writes a commitment direction into the residual stream, and the similarity of two questions' directions predicts how strongly their answers are coupled. How this direction is read, however, varies with model scale, instruction tuning and system prompt, and in one model the sign of the effect reverses.
WhyThe authors report that a model's own earlier answer writes a commitment direction shaping later answers, tested across 22 models and 3,600 question pairs.
In briefDhāra Bench: Evidence-Oriented Retrieval Benchmark for Indian Knowledge Systems
SummaryDhāra Bench is a retrieval benchmark for Indian Knowledge Systems, containing 325 questions in 12 categories and 60,121 text chunks from 15 source groups. Its categories were developed from 4,500 questions submitted to the Dhāra application, and its evaluation set combines corpus-aligned queries with expert-written questions. The paper reports results for a fixed BM25 baseline, including complete-evidence scores for multi-hop questions.
In briefMoSoLoco: a method for locating people in 3D from one camera using visual features and nearby people
SummaryThe paper presents MoSoLoco, a per-frame method that estimates the metric 3D position of people from a single monocular camera. It combines features from the pose detector with the detected 2D skeleton, without an extra RGB or depth encoder. The model can also use sparse information about other people in the same frame to refine each estimate. On JRDB, the best error is about 23 percent lower than the strongest keypoint-based baseline.
In briefACTOR: Time-Ordered Action Modeling for Literary Role-Playing AI
SummaryThis paper presents ACTOR, a framework that automatically builds character behavior trajectories, including non-dialogue actions, from raw novel texts and orders them by story time. Each behavior is represented with a Thought-Action-Observation structure, and the authors create a dataset from 143 literary works with counterfactual behaviors inconsistent with character personas. Models trained with Regularized Preference Optimization on these pairs showed substantially improved accuracy in a four-way next-behavior selection task across 17,560 evaluation instances.
In briefAutonomous agent that tunes MSA construction to improve protein complex structure prediction
SummaryThe paper studies how the depth and cross-chain pairing of multiple sequence alignments (MSAs) affect AlphaFold 3's prediction of protein complex structures, finding the best strategy depends on the target. The authors introduce an autonomous agent that treats a frozen AF3 as a black box, forming biological hypotheses and iteratively reorganizing MSAs and pairing choices, including deliberate non-pairing, using label-free confidence signals. Across 1,015 protein-protein interfaces from three benchmarks, the agent improves overall DockQ over the default pipeline and fixed recipes while using on average 60% fewer MSA sequences. The largest gains occur for inter-species heteromers, with mean DockQ rising from 0.465 to 0.612 on FoldBench and from 0.482 to 0.562 on HD439.
WhyThe authors treat MSA design as target-specific and adapt it at test time with an autonomous agent; they report DockQ gains over AF3's default on 1,015 interfaces across three benchmarks.
In briefEvery Clause Counts: Reranking Search Results to Follow Instructions
SummaryThe paper introduces DMReranker, a zero-shot reranker that splits an instructed search query into a topic and typed clauses, has a small judge model answer one narrow question per clause, and multiplies the probabilities into one score. On the InfoSearch benchmark it reaches a macro WISE of 50.9 and original-mode nDCG@10 of 82.9. It trails GPT-4o on the stricter SICR metric (19.2 versus 32.1), which the authors attribute to how products of probabilities behave when raw scores are compared across runs.
In briefCoworkOS: A framework for scaling AI agent coworkers to automate real-world tasks
SummaryThis paper presents CoworkOS, a general-purpose framework for long-horizon work that separates task strategy from execution management. The model decides actions, delegation, and acceptance, while the runtime maintains state and feedback. Humans and agents can edit the same documents, presentations, CAD models, and Blender scenes, and hierarchical workspaces and persistent memory support continuity across sessions. The authors report that CoworkOS is highly competitive with frontier agent applications such as Codex and DeepSeek Harness across 12 benchmarks and 6 categories of real-world applications.
In briefViAdminShift: A Vietnamese benchmark testing whether language models update knowledge after the 2025 administrative reform
SummaryThis paper presents ViAdminShift, a Vietnamese benchmark built around the July 1, 2025 administrative reorganization that cut Vietnam's provinces and communes. It includes 7,995 benchmark items and 2,281 retrieval passages. The authors test 67 models and find that many still give old provinces for old addresses, with the answer changing when a date is given in the prompt.
WhyThe authors build an event-level benchmark from Vietnam's 2025 administrative reform and report probing results on 67 models, including outdated-answer rates.
In briefA position paper arguing causal discovery needs explicit standards for judging its assumptions
SummaryThis position paper argues that causal discovery algorithms lack shared principles for choosing between them. It contends that identifiability assumptions are often chosen by convention, and that consistency results and benchmarks also rest on hard-to-test assumptions. The authors call for explicit standards for discussing these assumptions, with benchmarks stress-tested and ablated, so theoretical and empirical results can be combined.
Harald Kugler, Leonard Henckel, Sebastian Weichwald
In briefLaBraM at Scale: Expanding Data and Models for EEG Pretraining
SummaryThis paper extends LaBraM, an early EEG foundation model, by scaling up its pretraining data and model capacity while keeping its core encoder and objective. The authors train ten encoders ranging from 5.8 million to 1.97 billion parameters on three EEG collections. Across 23 downstream tasks, the best full fine-tuning result reaches 59.05% mean balanced accuracy, 3.61 points above the strongest baseline.
In briefBeyond Surface Metrics: Multi-Condition Requirements in Maintenance and AI Coding Benchmarks
SummaryThis paper examines how software maintenance tasks must satisfy multiple operational conditions, some with directional dependencies. Using a fixed 18-criterion taxonomy, it analyzes 2,925 Apache Fineract issues and 944 tasks from SWE-bench Pro, SWE-PolyBench Java, and SWE-EVO. Among benchmark task pairs matched on issue type, modified-file count, and code churn, most have disjoint fulfillment-condition sets, and time and numerical conditions relate differently across the two corpora. The authors recommend reporting condition breadth, co-occurrence, and directional relationships separately, and note that corpus comparisons are descriptive only.
In briefToward knowledge-centric AI: a paper on storing and reusing explicit knowledge across AI agents and generations.
SummaryThe paper argues that future AI should be built around an explicit, continuously evolving knowledge layer that is independent of any single agent. It describes a loop in which agents discover knowledge, the layer internalizes it, and agents across systems and generations use and revise it. The authors identify internalization as the main bottleneck and call for knowledge foundation models and a research agenda.
In briefAnalyzing Geometric Constraints in Grokking: Acceleration That Depends on Component Interactions in Modular Arithmetic
SummaryThis paper studies how spherical transformer configurations accelerate grokking, a delayed-generalization phenomenon, on modular arithmetic tasks. Using 960 controlled runs, the authors conducted a full factorial analysis of residual projection, readout weight normalization, logit scale, and weight decay. They report that the acceleration depends on interactions among these components and does not hold uniformly across all tested tasks and depths.
In briefPaddleMaterials: An End-to-End Deep Learning Toolkit for Materials Science
SummaryThis paper presents PaddleMaterials, an open-source framework that unifies data processing, training, evaluation, pretrained-model management, and workflows for six materials-learning task classes. These include property prediction, structure generation, machine-learning interatomic potentials, electronic-structure fields, spectrum elucidation, and microscopy image processing. A shared compiler backend lets users switch from dynamic-graph to compiled execution without changing model definitions. The authors report speedups across 65 model instances in 11 groups, and case studies on LiMnO2 and toluene compare model outputs against same-geometry DFT references.
In briefFrom PID to RL in HVAC Control: A Review of Control Architectures
SummaryThis review examines control architectures for building HVAC systems, which must balance energy use, occupant comfort, and equipment operability. It covers PID regulation at the thermostat level and formulates building control with reinforcement learning as a Markov decision process. The central contribution is a taxonomy of four ways RL can meet PID: tuning an existing loop's gains, directly replacing the regulatory controller, hybridizing with model predictive control, and generating supervisory setpoints and schedules. The authors argue these are nested attachment points on one HVAC plant rather than competing philosophies, and that deployment trade-offs depend on the specific control interface and required scaffolding.
Erfan Haghighat Damavandi, Davide Papurello, Mahdi Alibeigi et al.
In briefFlowBatt: Flow Matching for Probabilistic Battery Degradation Prediction
SummaryThis paper presents FlowBatt, a conditional generative framework that applies flow matching with a diffusion transformer backbone to model battery degradation trajectories. Conditioned on early-cycle capacity data, it generates full state-of-health (SOH) trajectories as probabilistic processes, and uses explainable AI analysis to improve the encoder that converts capacity matrices into conditioning vectors. On five public benchmark settings, FlowBatt achieves the best trajectory-wise SOH prediction performance, with errors below 2.5%, and is competitive on remaining useful life (RUL) prediction. By generating multiple plausible trajectories, it provides empirical prediction intervals.
Hamidreza Eivazi, Eslam Sharaawy, Stefan Wittek et al.
In briefControlling "sharpening" in LLM post-training via k-based pressure
SummaryThis paper studies sharpening, where reinforcement learning post-training of LLMs improves accuracy but reduces behavioral coverage, and examines why repeated-sampling objectives like pass@k and max@k mitigate it. The authors derive closed-form optimal policies for KL-regularized pass@k and max@k objectives and show that standard @1 objectives can reach the same optima through prompt-adaptive regularization or reward transformation. They introduce KAPPO, which adapts existing @1 methods to @k with a one-time reference calibration, and report experiments on math reasoning and code generation showing direct control over sharpening with single-completion training matching multi-completion optimizers.
Simon Matrenok, Nicolas Menet, Nicolas Flammarion et al.
In briefTurning Building Images into Editable Procedural 3D Models
SummaryThis paper studies reconstructing an executable procedural description of a building from in-the-wild images, so it can be instantiated as an editable 3D asset. The method works in two stages: a Global Structure Layout Tree recovers facade-level split-and-repeat organization, and a Component Parameterization schema maps representative facade tiles to procedural attributes. Both are deterministically compiled into executable CGA programs. Experiments report structural consistency and architectural detail on regular and challenging images.
In briefDepthGraft: Structural Regularization via Hierarchical Cross-Layer KV Reconstruction
SummaryThis paper addresses depth-wise information dilution in Pre-LN large language models by treating cross-layer KV sharing as a structural regularizer. It proposes DepthGraft, which reconstructs deep-layer KV from stride-aligned shallow-middle source pairs. Experiments on dense and MoE models from 0.5B to 16B parameters show downstream gains under both GQA and MLA, including a 7.0% absolute MMLU-Pro improvement on the 16B MoE model.
In briefAligning diffusion model optimization with what people actually intend
SummaryThis paper examines how diffusion models are optimized mainly for measurable goals such as aesthetics, which may not match what users truly want. The authors systematically analyze how human intent is expressed, turned into optimizable objectives, applied to model behavior, and evaluated through outputs. They also discuss trade-offs in aligning with intent in personalized image generation and gaps between automated metrics and real user judgments.
In briefKlaskTron: An open-source, low-cost robot platform for adversarial multi-agent reinforcement learning
SummaryThe paper presents KlaskTron, an open-source, desktop-scale robotic testbed costing under $2,500, built around the tabletop game KLASK for adversarial multi-agent reinforcement learning. It uses dual CoreXY gantries with high-torque brushless motors and includes hardware designs, an NVIDIA Isaac Lab digital twin, and a validated sim-to-real baseline. The authors report that trained agents transfer zero-shot to the physical system and exhibit emergent adversarial behaviors.
Aswin Ramachandran, Jona Schulz, Maurus Derungs et al.
In briefZnaki: A Russian Fingerspelling Recognition Dataset
SummaryThis paper introduces Znaki, the first public dataset for Russian Sign Language fingerspelling. It contains 37,252 videos (35.4 hours) of 1,593 proper-name sequences recorded by 68 native signers and validated by experts. The dataset covers all 33 Cyrillic handshapes. Two baseline models adapted from American Sign Language recognition are reported, reaching character-level accuracy of 0.932 and 0.886.
Petr Surovtsev, Pavel Korotaev, Alexander Nagaev et al.
In briefAdapting Teachers to Student Reasoning Prefixes for On-Policy Distillation
SummaryThis paper studies a weakness of on-policy distillation, where a teacher model becomes less reliable when supervising student-generated reasoning prefixes, especially longer ones. The authors propose Student-Prefix Teacher Adaptation (SPTA), which trains the teacher with reinforcement learning with verifiable rewards to continue from collected student prefixes before distillation. Experiments on mathematical reasoning benchmarks report that SPTA improves average student performance over standard OPD and OPD with task-level RLVR-adapted teachers.
In briefBehavioral affinity method for zero-shot image classification with unlabeled batches
SummaryThis paper introduces Behavioral Affinity Transduction (BAT), a method for classifying images in small unlabeled batches with zero-shot vision-language models. BAT builds a sparse graph from each image's scores over the class vocabulary and uses these neighbor links during prediction. Across diverse datasets, BAT improves the mean accuracy of every tested readout. In the main RN50 comparison, soft and hard EM-Dirichlet gain 4.22 and 2.72 percentage points.
In briefCUA-Master: Scaling computer-use agents via open-world recursive improvement
SummaryThis paper presents CUA-Master, a nested double-loop framework for recursively improving computer-use agents (CUAs). The outer loop continuously incorporates open-world data and uses the model's weaknesses to guide data construction and optimization, while the inner loop mines workflows from instructional videos as seed tasks and expands them into diverse, verifiable tasks. The authors report that CUA-Master-27B scores 80.2 on OSWorld Verified, 36.9 on OSWorld 2.0, 43.9 on WeaveBench, and 25.0 on Agents' Last Exam, with continued gains across iterations.
In briefSurgPace: Difficulty-Aware Progressive Distillation for Compact Surgical AI Models
SummaryThis paper presents SurgPace, a framework for progressively distilling large surgical vision-language models into smaller, more efficient ones. It estimates how difficult each training sample is using prediction confidence and Lorentz geometric consistency, and trains from easy to hard samples. It also aligns student and teacher representations at feature, logit, and probability levels while shifting focus from imitating the teacher to task supervision. Tested on the Surg-396K benchmark, SurgPace outperforms strong multimodal distillation baselines.
In briefComplianceAtlas: An Agentic AI Framework for Mapping and Fixing EU AI Act Compliance Gaps
SummaryThe paper presents a taxonomy of technical compliance gaps in areas such as accuracy, robustness, logging, and human oversight, drawn from the literature. It introduces GapToCompliance, a two-stage framework with one agent that extracts gap types hierarchically and another that ranks remediation actions from a catalog. Evaluated on 9 AI systems across 6 AI families, it achieves a gap-type F1 of 0.831 and a remediation P@5 of 0.872, outperforming the baselines.
In briefKAUSTics-Net: Diffusion-based restoration of through-water UAV images with a semi-synthetic dataset
SummaryThis paper addresses distortions in shallow-water UAV imagery caused by refraction, caustics, and wavelength-dependent attenuation. The authors build KAUSTics, a semi-synthetic dataset of 1,280 reference patches and 12,800 degraded observations. They propose KAUSTics-Net, a pixel-space conditional diffusion framework, and compare it against eight restoration methods with ablations and cross-dataset experiments.
Dario Scilla, Omar A. Lopez, Kasper Johansen et al.
In briefInner optimizers in active-learning Kriging reliability analysis: comparison and underestimation risks
SummaryThis paper compares five budget-matched continuous optimizers with candidate-pool methods inside active-learning Kriging reliability analysis across eleven problems. The pre-specified hypothesis that the hybrid swarm method with L-BFGS-B polishing would outperform was not supported. In many runs an unlearned failure mode led to underestimated failure probability, and the error criterion accepted some substantially underestimated results. The authors recommend default pool selection and coverage checks.
In briefFrom Local Skills to Long-Horizon Web Tasks: A Progressive Skill Exploration Method for LLM Web Agents
SummaryThis paper addresses the difficulty of training LLM-based Web agents, where successful trajectories are rare and sparse terminal rewards provide little feedback. It proposes SCOPE, a framework that iteratively discovers reusable local skills, composes them into skill chains, and expands toward new subskills guided by failure experiences. The learned skill chains are internalized into policy parameters through agentic on-policy distillation. Experiments on WebArena, WebVoyager, and WebShop show 3.9% to 18.6% relative gains over the strongest baselines.
In brief3D Gaussian Splatting with Neural View-Dependent Color Fields: a method to improve 3D scene geometry in novel view synthesis.
Summary3D Gaussian Splatting (3DGS) is a widely used tool for novel view synthesis, but its discrete Gaussians often yield poor depth, few geometric details, and many floaters. The paper learns the radiance field within a continuous view-dependent color field parameterized by neural networks, adds a constraint improving the Lipschitz continuity of the color field, and introduces self-learning strategies that need no external priors. Comparisons on widely used benchmarks are reported to show better high-fidelity geometry recovery.
In briefHow Fine-Tuning Data Preprocessing Affects LLM Recommendation Utility and Demographic Fairness
SummaryThis paper examines how preprocessing choices in fine-tuning data affect an LLM-based generative recommender's accuracy and demographic-related fairness. The authors combine supervised fine-tuning, catalog-constrained decoding, and prompt augmentation, then compare item-title representation, demographic balancing, and item-knowledge enrichment on MovieLens-1M. Using original titles raised HR@10 from 10.67% to 12.47% and Coverage@10 from 41.36% to 47.88%. Age balancing and knowledge enrichment narrowed some group-level utility gaps, while gender balancing reduced utility without improving the gender-related indicators evaluated.
In briefPredicting lung adenocarcinoma invasiveness from 3D CT using dual-granularity multimodal prompt learning
SummaryThis paper proposes a multimodal prompt-learning framework that pairs 3D CT images with radiological text to distinguish invasive adenocarcinoma (IAD) from pre-invasive lesions (Pre-IAD). It aligns image and text at both whole-volume and local vascular-keyword levels, and requires only CT images at inference. On 1,593 pathologically confirmed nodules, the method reached an AUC of 0.9499 and 87.92% accuracy on the test set, exceeding an image-only baseline with the same visual backbone (AUC 0.9297).
In briefDiagonal equivalence and retrieval rank in state space models
SummaryThe paper classifies real and complex state space models with linear transitions, separating fixed-basis diagonal equivalence from input-output equivalence. It introduces HybridMamba, which combines features of Mamba 1, 2, and 3, and reports that it outperforms Mamba 1 through 3 on the hardest tier of small-model multi-token recall experiments. In single-seed language models of about 500M parameters, both proposed designs improve test loss over Mamba 1 and 2 at 2B tokens, but the routing gains do not lower language-model loss.
In briefLUCA: A Synthetic World for Testing Experimental Reasoning in AI Agents
SummaryThis paper introduces LUCA, a configurable synthetic environment with fully known ground truth for evaluating how AI agents reason through experiments. Agents learn unknown regulatory systems under a limited experimental budget, then predict unseen interventions, diagnose hidden faults, and attempt rescues. Tests on seven frontier LLMs and three baselines show these abilities vary separately, with diagnosis differing about 7x across models versus 1.4x for prediction. Agents often rescue malfunctioning systems while diagnosing the wrong fault.
In briefRay-SCR: Making Scene Coordinate Regression Work Across Different Camera Types
SummaryScene coordinate regression (SCR) is a visual localization method, but it generalizes poorly across very different cameras such as omnidirectional or strongly rectified imagery. The paper introduces Ray-SCR, a lightweight extension to existing SCR architectures that conditions local features on calibrated viewing rays and refines them with a shallow local Transformer. Using sparse SfM points as metric anchors, a single scene representation built from a panoramic sequence can localize pinhole, fisheye, and omnidirectional queries, achieving state-of-the-art cross-camera performance.
In briefDR Pixel: Deep Research Directly on Document Page Images with Vision-Language Models
SummaryThis paper introduces DR Pixel, an open-source multimodal research agent that retrieves and reasons over rendered page images instead of OCR-extracted text, across a 1.06M-page corpus. The authors also build DRPixelBench, a benchmark of scientific papers requiring reasoning over text, figures, tables, and equations. The approach outperforms OCR-based baselines even without training, and a counterfactual credit reweighting method further improves retrieval recall and accuracy.
WhyThe authors propose a page-native multimodal research agent with counterfactual credit reweighting, reporting gains over OCR baselines on DRPixelBench and four other benchmarks.
Tiansheng Hu, Yichuan Wang, James V. Talwar et al.
In briefA fine-grained difficulty ladder and benchmark for how language models track changing states
SummaryThe paper proposes a four-level hierarchy of word problems, from simple immutable variable assignment to general finite state tracking, to evaluate how well language models track state. It theoretically compares transformers with recurrent architectures such as Gated DeltaNet, and builds StateBench, a long-horizon benchmark based on these levels. Direct training experiments largely match the theoretical predictions, while pretrained models, both open and closed, do not robustly solve even soluble state tracking.
WhyThe authors propose a four-level state-tracking hierarchy and benchmark, with theoretical characterizations of several architectures and direct-training and pretrained-model evaluations.
In briefTesting Hybrid Retrieval, NLI and Open-Source LLMs for Verifying Scientific Claims
SummaryThis study examines whether a claim is supported, contradicted, or lacks enough evidence in scientific text. The authors retrieve the top five relevant sentences per claim using hybrid semantic and BM25 search, then compare an NLI baseline with Llama 3.1 8B, Qwen3 8B and Mistral 7B. Qwen3 achieves the highest Macro-F1 at 0.7937, and a conservative fusion of NLI and LLM predictions lowers performance rather than improving it.
Lakshmi Narayanan P, Narenkarthik K, Naren Kumar C
In briefAvatar Intelligence: Persistent AI Agents That Extend Human Decision-Making
SummaryThis paper proposes Avatar Intelligence, a paradigm in which persistent personal AI agents represent users' goals, values, and constraints and make decisions on their behalf across contexts and time. It outlines a maturity roadmap from human-directed execution to autonomous operation within delegated domains. The authors introduce the Agentic Oversight Index (AOI) to evaluate delivered value against agent operating cost and human oversight burden, and identify three research challenges.
In briefMIRA: A training method that makes multimodal AI models more robust to misleading visual information
SummaryThis paper identifies a "perception-reasoning gap" in existing reinforcement learning methods for multimodal reasoning, where small visual errors compound into hallucinations. The authors propose MIRA, which injects filtered deceptive visual contexts during training as challenging traps and uses a group-wise reflection mechanism so the model learns to detect and correct inconsistencies between visual cues and logical constraints. Experiments report an average 7% gain on in-domain benchmarks and consistent improvements on vision-centric benchmarks, without human annotations.
In briefVL-Dedup: Graded Multimodal Matching for Image-Text Deduplication
SummaryThis paper argues that image-text datasets contain duplicates, varied captions, reused captions, and mismatches that need different handling, and reframes deduplication as graded multimodal entity resolution evaluated at the pair, cluster, and retained-set levels. The proposed VL-Dedup method groups records of the same visual entity into size-capped clusters, where a vision-language model and CLIPScore decide which captions to keep or flag. The authors also release VL-DedupBench, containing 50,000 COCO and SOP records and 66,788 labeled pairs.
In briefNetwork fusing spectral and spatial data to classify multi-source remote sensing images
SummaryThis paper presents CRFNet, a lightweight dual-branch network for classifying multi-source remote sensing images by combining spectral and structural information. It includes a cross-layer attention module that propagates selected attention relations across layers to strengthen discriminative spatial dependencies. A frequency-domain fusion module combines frequency interactions and cross-modal channel attention, then converts the result into spatial gates to modulate complementary features. The authors report that CRFNet outperforms several state-of-the-art methods on the Houston 2013 and Augsburg datasets.
In briefA network that jointly classifies hyperspectral and radar/LiDAR remote sensing data
SummaryThis paper proposes SPCNet, a Shared-Pattern Collaborative Network for jointly classifying hyperspectral (HSI) and SAR/LiDAR remote sensing data. Its Shared-Pattern Collaborative Block combines a shared large-kernel operator with a dynamic convolution branch to capture common spatial structures and content-adaptive local features. A Region-Adaptive Modality Allocator assigns spatially varying modality weights for feature fusion. Experiments on the Houston2013, Augsburg, and Houston2018 datasets show better classification performance than several state-of-the-art methods.
In briefPACERec: A recommendation method that adapts LLM item embeddings progressively using collaborative behavior evidence
SummaryThe paper proposes PACERec, an ID-free progressive adaptation framework for sequential recommendation. It learns a shared velocity field from items with sufficient interactions using frequency-weighted flow matching, incorporating preceding interaction contexts into static LLM embeddings. The authors report that experiments on three real-world datasets and three sequential backbones show improvements over competitive baselines in both overall and long-tail recommendation.
In briefPlanContext: A Benchmark for Context-Aware Driving Decisions
SummaryThis paper introduces PlanContext, a benchmark for evaluating how self-driving systems make decisions that depend on context. It keeps the scene fixed while varying the high-level driving intention, creating paired evaluations across six driving contexts. Tests on 9,561 scene-condition instances from NAVSIM and 23 public end-to-end models show that high conventional driving-quality scores do not consistently mean strong route-level context-dependent decisions.
In briefArchitecture for a trustworthy, governable Internet of Agentics
SummaryThis perspective paper examines how LLM applications are shifting toward tool-using, multi-agent workflows and the interoperability needs this creates across frameworks, runtimes, and organizational domains. It treats heterogeneous agents as governable service participants within an Internet of Agentics view. The paper outlines a three-layer progression from a fragmented ecosystem to interoperable agents and then trustworthy, governable autonomy, covering protocols, systems abstractions, and links to services computing.
In briefNetworked AI: A Network of Evolving AI Agents
SummaryThis paper proposes Networked AI, a system made of interacting, heterogeneous, and continuously evolving AI agents. It sets out a three-layer architecture (Agent, Interaction, and Regulation layers) and a unified theoretical framework drawing on data science, signal processing, and system control. The authors report that resource demand can scale near-linearly, in contrast to exponential growth in standalone end-to-end systems. Applications discussed include IoT, embodied and spatial intelligence, industrial intelligence, and cellular infrastructure.
In briefMuCHMDPoser: Multi-camera RGB full-body tracking for VR avatars
SummaryFull-body tracking in virtual reality is poorly constrained when using only consumer headsets and controllers, which leaves lower-body pose poorly estimated. This work adds external multiview RGB cameras and integrates RGB-based pose estimation to reduce lower-body tracking errors. It also evaluates a one-time headset-to-camera calibration that needs no physical markers. The authors implement the framework, which runs end-to-end animation at 60 FPS on an Nvidia RTX3080 GPU.
Antoine Maiorca, George Fletcher, Thierry Ravet et al.
In briefSAGE: A linguistic method for spotting AI-written peer reviews
SummaryThe paper proposes detecting AI-generated peer reviews by measuring subjectivity, a trait typically human, at semantic and pragmatic levels. A lightweight classifier using only these linguistic features reaches an AUROC of 0.9566. Combined with the existing Sem-Detect detector, it raises Sem-Detect's TPR at 1% FPR on the APRDB benchmark from 0.874 to 0.985.
Gonzalo Freijedo Aduna (IJN), Charles Yaiche (CRIStAL), Patrick Bas (CRIStAL) et al.
In briefPRIME: a feedback-to-revision routing layer for scientific idea generation
SummaryThis paper introduces PRIME, a plug-and-play decision layer that turns feedback on LLM-generated scientific ideas into revision plans for existing ideators. It works in three stages: structural attribution of feedback, counterfactual routing that compares projected outcomes of alternative revision scopes, and compilation of the chosen intervention into an ordered generator plan. Validated on 112 human-annotated idea-feedback pairs, PRIME improves successor quality and routing reliability, and removing counterfactual routing markedly lowers quality.
In briefRobust cell-free MIMO integrated sensing and communication under uncertainty via a graph neural network
SummaryThis paper addresses beamforming for integrated sensing and communication in cell-free MIMO systems, where channel state information errors and clutter create reliability challenges. The authors propose an impairment-aware heterogeneous graph neural network that takes the CSI error level and clutter covariance as edge features, so one model can handle both impairments. Simulations show it outperforms baselines in sensing-only and communication-only settings and generalizes to unseen error levels and clutter counts, making the system more resilient.
Oways Alsoloh, Robert-Jeron Reifert, Aymen Khaleel et al.
In briefFinding the Most Deviating Subgroups at Any Support Level
SummaryThis paper studies subgroup discovery, which seeks regions of the feature space where a target variable deviates from its overall distribution. It gives the first theoretical and structural characterization of subgroups with maximum KL divergence under arbitrary support constraints. The authors show that every point on the deviation-support Pareto front is a union of whole atoms plus at most one fractional atom. They propose a two-step method that learns a homogeneous partition and then selects atoms greedily, and report results on 19 benchmark datasets.
In briefExtracting Data from Criminal Court Rulings for Legal Research
SummaryThis paper presents a large language model pipeline that turns criminal verdicts into a structured dataset through attribute exploration, text span extraction, and value normalization, guided by annotation instructions that legal experts can revise. It is evaluated on driving under the influence and driving despite a ban. For the former, revising the instructions raised correctly extracted values from 52% to 70% and improved agreement between human annotators.
Tom\'a\v{s} Knap, Ivana Kvapil\'ikov\'a, Jan \v{C}ern\'y et al.
In briefVideoEditor-MoE: Scaling Complex Video Editing with Mixture of Experts
SummaryThis paper introduces VideoEditor-MoE, the first open-source mixture-of-experts framework for video editing, aimed at complex edits that fulfill multiple requests at once. Besides sparse MLP-MoE layers, it adds MoE-Attention, which expands attention capacity through routed key-value experts while keeping global interactions among video tokens. Initialized from a 5B video editing model, it scales to 14B and 56B parameters, with about 8B parameters activated per token in both variants.
In briefReinforcement learning for choosing the next measurement in single-pixel imaging
SummaryThis paper treats adaptive compressed sensing for single-pixel imaging as a Bayesian sequential optimal experimental design problem and reformulates it as reinforcement learning. The agent learns a policy that maps the current posterior belief about the image directly to the next binary illumination pattern, using expected information gain as the reward. The authors evaluate the scalability of the learned policy under two parameterizations, study performance under approximate posteriors and noise, and propose design constraints motivated by biomedical applications.
Marcos Antonio Obando, Tianyuan Wang, Tanja Tarvainen et al.
In briefVERS: A method that steers reasoning models by editing text explanations of their activations
SummaryThe paper introduces VERS, which converts edits to natural language explanations of a reasoning model's activations into one fixed direction and adds it at paragraph breaks to shorten generated reasoning. On GSM8K it reduces generated tokens by 45.8% with an accuracy change of -0.82 percentage points. On MATH-500 it saves 31.3% of tokens at -3.05 points. On competition problems, accuracy drops at every model size.
In briefTSB-SEG: A systematic benchmark for testing methods that split time-series data into segments.
SummaryThis paper introduces TSB-SEG, a unified benchmark for time-series segmentation, covering both change point detection and state detection. It includes 592 univariate and multivariate time series from 8 datasets and compares 26 algorithms. An open-source library, tsseg, is provided to ensure reproducibility, and the authors offer guidance on method selection based on data properties.
In briefWhen design thinking works: a convergence theorem for the iterate-test loop
SummaryThis paper formalizes the ideate, prototype, and test loop of design thinking as a search over a finite prototype space whose utility is learned from noisy evaluations. Under five stated conditions, the authors prove that the prototype a team believes best converges almost surely to a local optimum. Finite-horizon simulations across three landscape types show teams overestimate their chosen concept by 4-8%, and suggest building one to three prototypes per five tests while re-testing leaders.
Pawe{\l} Kura\'s, Adrian Michalski, Alicja Gerka et al.
In briefGPU-based real-time textured mesh reconstruction with seam-aware multi-view brightness correction
SummaryThis paper presents a GPU-native volumetric-video system that reconstructs graphics-ready textured triangle meshes in real time from multiple RGB-D sensors. Fusion, mesh extraction, texturing, inpainting, and seam correction all stay on the GPU, avoiding transfers of intermediate data back to the host. The method corrects exposure seams with a per-camera low-frequency gain field baked into the texture atlas. On the CMU Panoptic dataset, it is faster than the baselines at every tested resolution and sensor count, with peak speedups above tenfold over nvblox and Open3D, and its re-rendering quality approaches an offline textured-meshing baseline.
In briefFederated learning that fuses multiple data types when some are missing and sites differ
SummaryThis paper studies multimodal fusion in a hybrid federated setting where data are split across both samples and feature blocks. It proposes 2D-FedSMKL, a method combining availability-corrected modality scores, group-kernel regularization, and smooth worst-site risk aggregation. Curvature bounds, block-gradient equivalence, and a conditional accelerated convergence rate are derived for binary and multiclass scores. On simulations and two public activity-recognition benchmarks, worst-site aggregation improves worst-site performance under heterogeneity, and the selected model is not significantly worse than the strongest reference at the worst site.
In briefUsing a Vision-Language Model as a Judge for Person Re-Identification
SummaryThe paper uses a general-purpose vision-language model, applied zero-shot, to rerank short candidate lists in cloth-changing person re-identification. On the four CHIRLA scenarios it raises rank-1 accuracy in every scenario, averaging 12.9 points above the best expert model. To cut cost, the VLM is called only when the two expert models disagree at rank 1, which removes 40% of calls without a significant rank-1 drop.
WhyThe authors report that zero-shot VLM reranking improves rank-1 on all four CHIRLA scenarios, averaging 12.9 points over the best expert, with a disagreement-based gate cutting VLM calls by 40 percent
Bessie Dominguez-Dager, Miguel Cazorla, Vincent Lepetit
In briefBetter regret bounds for tabular reinforcement learning under local differential privacy
SummaryThis paper studies regret minimization in episodic tabular reinforcement learning with S states and A actions under non-interactive local differential privacy. For the single-step case, the authors establish a minimax regret bound and achieve it with an EXP3-based algorithm and a unary encoding mechanism. For multi-step horizons with small privacy parameter, they give matching-order lower and upper bounds that improve previously known results.
In briefVizDAGuide: Training Small Vision-Language Models to Answer Questions About Misleading Charts Using Graph-Structured Supervision
SummaryThe paper presents VizDAGuide, which represents intermediate reasoning states and their dependencies as a directed acyclic graph and assigns supervision weights by role. The graph is serialized into a single training target for misleading-chart question answering in both open-ended and multiple-choice formats. Experiments on two benchmarks with Qwen3-VL and InternVL3.5 report gains over a content- and weight-matched linear chain-of-thought baseline. Ablations indicate that dependency structure and weighted supervision work best when combined.
In briefMulti-agent AI: when paired senders beat solo ones
SummaryThis paper studies classifiers that each see different parts of one object and send short messages to a receiver that identifies its class. On three digit datasets, messages designed jointly outperformed messages designed separately by 18.07 to 22.68 percentage points on average with two-bit messages. The authors argue that individual statistics alone cannot show whether a partnership helps, so cooperation must be measured in pairs.
In briefEfficient Hybrid Distillation: Combining Score and Adversarial Objectives for One-Step Diffusion
SummaryThis paper studies how to combine score distillation and adversarial distillation to speed up diffusion models. Score distillation works well at high noise levels but lacks fine supervision at low noise, while adversarial distillation is the reverse. The authors propose Efficient Adversarial Score Distillation (EASD), which uses regime-adaptive sampling so each objective focuses on its effective noise interval, and apply it to DiT-based Flow Matching models. The paper reports a one-step FID of 1.39 on ImageNet 256×256 with the SiT-XL/2+REPA model.
In briefComparing one shared model against per-ship fuel prediction methods for pre-departure use
SummaryThis paper compares 11 methods for predicting ship fuel oil consumption before departure, across 21 vessels, using identical inputs available before departure and testing on voyages after each vessel's training records. A global model that describes each vessel by its identity and seven ship specifications performs best in every setting examined. For in-fleet vessels, Global LightGBM reaches an overall MAPE of 8.92%, versus 11.14% for Per-Ship LightGBM. The authors conclude that per-ship and vessel-specific adaptation models are unnecessary, since one global model can serve a whole fleet.
In briefQ-CamTuning: A Multimodal Language Model That Reasons About Fine Image Quality to Auto-Tune Livestream Cameras
SummaryThis paper presents Q-CamTuning, a multimodal large language model that automatically tunes cameras for livestreaming. The authors built FGLive-Comments, an instruction-tuning dataset with annotations on fine-grained image pair preferences and quality reasoning based on global and local visual evidence. Experiments show Q-CamTuning outperforms existing image quality assessment models in quality scoring and fine-grained pair ranking, and a user study confirms its camera tuning suggestions.
In briefBLPO: A Policy Optimization Method for Long-Horizon LLM Agents
SummaryThis paper proposes Beyond-Local Policy Optimization (BLPO), a reinforcement learning method for training large language model agents on long-horizon tasks that require coordinated multi-step actions. It estimates task-progress advantages from experience shared across tasks and training updates, using a global memory with reliability weighting to complement local comparisons. Experiments on ALFWorld and WebShop with 1.5B and 7B policies report higher task success than the evaluated baselines under matched training budgets, along with faster learning and shorter trajectories.
In briefILLRAW: Illumination-aware object detection directly from RAW camera images
SummaryExisting RAW-based detectors usually convert RAW inputs into sRGB-like images with learnable processing modules before feeding off-the-shelf detectors, separating representation learning from detection. This paper proposes ILLRAW, an illumination-aware end-to-end RAW object detection framework that inserts a lightweight feature adapter at the detector input to learn detection-oriented RAW features directly. It estimates an illumination mask, models Bayer-channel responses with channel-sensitive embeddings, and fuses them with an illumination attention fusion module. Experiments show clear gains on the AODRaw benchmark and competitive results on LOD and PascalRAW, especially under dark exposure.
In briefFinding Errors Is Not Enough: Selective Escalation to Other Models for Sentiment Analysis
SummaryThis paper studies how to improve a document-level entity-level sentiment classifier under limited resources, mainly using South Slavic news. It compares data acquisition orders as training sets grow and uses uncertainty to decide which predictions to revisit and whether an alternative encoder or local LLM should correct them. Two routing policies chosen on development data each send 30% of records to an encoder-assisted LLM, yielding gains of 1.02 Macro-F1 points on SentiCoref and 1.71 on NewsMTSC, at substantially different inference costs.
Nishan Chatterjee, Matthew Purver, Antoine Doucet et al.
In briefTraining a BERT model to rate literary sentiment using labels from a large language model
SummaryHuman-annotated literary text is scarce, which limits sentiment analysis tools for literature. This study tests whether labels generated by a large language model can train smaller models. The authors fine-tuned five BERT models with ordinal regression on 10,889 literary sentences labeled by Gemma 4 31B on a ten-point scale. The models were evaluated on 1,923 human-annotated sentences from Hemingway's The Old Man and the Sea, where BERT-large reached a Kendall's τ of 0.589, close to the teacher model's 0.604.
In briefClassification Aggregation with Multiple Memberships: combining several people's category choices into results where an object may belong to more than one category.
SummaryThe paper studies how to aggregate individuals' classifications of objects when each person assigns one category per object but the combined result may place an object in several categories. Under the natural settings of Independence and Supportiveness, three new axioms (Responsibility, Split Stability and Rejection Stability) each characterise junta rules, which assign a category to an object exactly when some member of a fixed coalition does so. The paper also examines an infinite population setting to clarify differences among the axioms.
In briefUtility-aware generative recommendation with utility-enriched codes and guided decoding
SummaryThis paper proposes UAGR, a method for generative recommendation that incorporates downstream multi-objective utility into both semantic identifier construction and decoding. It aims to keep high-utility candidates from being pruned prematurely during beam search. According to the paper, offline experiments and online A/B tests show its effectiveness and deployability, and it has been fully deployed on Kuaishou, increasing video watch time by 0.206% online.
In briefCAREFUL-Bench: A Judge-Free Benchmark That Scores the Side Effects of Multi-Turn LLM Function Calls
SummaryThe paper argues that the Berkeley Function Calling Leaderboard's multi-turn checker judges only final state and ignores side effects of intermediate calls, so models that ask for confirmation before destructive actions are penalized. The authors infer effect declarations for all 128 functions and add a judge-free replay scorer and a confirmation-aware protocol with a deterministic simulated user. Across 11 open models, effect-aware scoring keeps the official ranking while exposing differences in execution quality and in handling of confirmation.
In briefSHq: Learnable Quantization Grids for Compressing LLM Weights
SummaryThis paper introduces SHq, a closed-form family of lookup-table grids for block-scaled low-bit quantization of large language model weights. Its two parameters, a companding strength and an offset, adjust level spacing and asymmetry, with the offset letting the unused FP8 scale sign bit select a mirrored grid. Tested on five models from 3B to 35B parameters at 2 to 6 bits, SHq gives the lowest KL divergence in every 3- and 4-bit setting and reduces NVFP4's KL divergence by 41-67% at 4.5 bits per weight.
In briefOne Image Is Enough: Self-Generated Supervision for Personalizing Human Video Diffusion
SummaryThis paper addresses pose-driven human video diffusion models, which struggle to keep a person's identity consistent when the target pose differs from the reference image. Instead of requiring multi-pose subject videos, the authors synthesize pseudo-ground-truth frames at key poses with a pose-conditioned generator, select the most identity-consistent ones using a reference-only sampler, and train a per-subject low-rank adapter on them. On TikTok and UBC Fashion, the adapted model improves over the base on every metric and matches or surpasses adaptation using five real frames.
WhyThe authors replace real multi-pose subject video with synthesized pseudo-ground-truth frames for personalization, reporting gains on TikTok and UBC Fashion.
Sadman Rakib Pinon, Adamo Orsini, Yanis Falaki et al.
In briefSplitting Prediction Uncertainty by Input Specification Gaps
SummaryThe paper notes that classical predictive uncertainty decompositions assume a fixed, complete input and ignore how much missing information could affect predictions. The authors formalize this setting and decompose prediction risk into specification, residual aleatoric, and epistemic uncertainty. Experiments indicate that specification uncertainty best detects underspecified inputs, while residual aleatoric uncertainty best detects in-distribution prediction errors.
Roman Vashurin, Mikhail Seleznyov, Assel Yermekova et al.
In briefGenerative diffusion on long-range correlated landscapes: how the reverse process commits to data
SummaryThis paper studies why generative diffusion works, using the Gibbs measure of a high-dimensional Gaussian random potential with long-range correlations as an analytically tractable data model. With the exact score, the authors show that the reverse process gradually resolves the branches of pure states and commits to a single one within a speciation window. They describe the score as a foam of slab-like cells separated by shock-like boundaries, and show that the denoising error drops in stages that follow this process.
WhyThe authors analyze reverse diffusion on Gibbs measures of long-range correlated Gaussian potentials, introducing a foam structure and Larkin-type argument for O(1) energy gaps.
Jerome Garnier-Brun, Giulio Biroli, Marc M\'ezard et al.
In briefBalancing Self-Supervised Learning by Finding and Generating Rare Data
SummaryThis paper introduces BRIDGE, a label-free method for improving self-supervised learning on long-tailed unlabeled data. The current encoder scores images by their mean kNN distance, and a budgeted rule selects images where support is thin, after which a frozen image-to-image diffusion model generates local variants of them for continued pretraining. Across five long-tailed, web-scraped and synthetic source settings, BRIDGE achieves the highest seven-task linear-probe and kNN average among the compared methods.
Filipe Laitenberger, Hirokatsu Kataoka, Yuki Asano et al.
In briefVerifying Agent Actions by Trying to Refute Them Before Execution
SummaryThe paper introduces BlackSwan, a procedure in which language model agents predict the intended effect of each state-changing action, form testable hypotheses about how it could be wrong, and test them with existing observations or non-mutating actions before executing it. It executes the action only if no hypothesis survives testing. On AppWorld tasks, BlackSwan improves success rate over the strongest baseline by 14.9 and 25.2 percentage points with and without thinking mode. It also improves WebShop performance by 12.7 percentage points.
In briefExact fusion of softmax attention heads may need exponentially many atoms
SummaryThe paper studies how many key-value entries, or atoms, are needed to exactly preserve the output when several independently normalized attention heads are fused into one. For a specific function it proves an exponential lower bound that is tight when the head count divides the dimension. It also covers exact preservation for pretrained-model caches and reports learned-cache comparisons on two disjoint banks of 64 contexts.
