In briefWhat social media teaches about slow-building AI risks
SummaryThis position paper argues that AI safety should address harms that accumulate gradually through everyday, repeated use, not only vivid catastrophic events. Drawing on social media research, it examines how narrowed perspectives, weakened independent judgment, and growing user profiles might build up in conversational AI. Sycophancy, manipulation, and privacy loss are presented as a case study, and the authors call for evaluating cumulative patterns of use rather than assuming repeated use is harmful.
In briefTriSAR-3D and PaintFusion: A Trimodal LiDAR-RGB-Thermal Dataset and Method for Person Detection in Search and Rescue
SummaryThis paper notes that most multimodal 3D person detection research targets automotive scenes, leaving unclear how well it transfers to rescue settings with lying, occluded, or unstructured-environment people. The authors introduce TriSAR-3D, a dataset recorded with a 128-channel LiDAR, RGB camera, and thermal camera covering varied poses, indoor and outdoor locations, and different times of day. They also propose PaintFusion, a trimodal pipeline fusing PointPainting evidence with a simple MLP, reporting 65.4 AP@50 and 51.6 ms per-frame latency. Removing lying instances from training collapses lying-pose AP@50 to near zero across detectors, indicating that dataset diversity strongly affects performance.
In briefRealCLI: A Unified Benchmark for Evaluating Human-Agent Collaboration in CLI Agents
SummaryThis paper introduces RealCLI, a paired benchmark that evaluates both task execution and human-agent collaboration quality on realistic command-line interface tasks. It offers two modes: REALCLI-EXEC, which provides full specifications upfront, and REALCLI-COLLAB, where a scripted user supplies information, corrections, and authorization decisions progressively. The best frontier configuration averages only 49.2%, and its peak joint collaboration score is 37.6%, with models struggling in steerability, trust calibration, and safety boundary management.
In briefKFS⁴LAM: adaptive keyframe selection and structure-aware mapping for endoscopic Gaussian splatting SLAM
SummaryThis paper presents KFS⁴LAM, a Gaussian-splatting SLAM method for endoscopic scenes such as colonoscopy. It combines motion-guided adaptive keyframe selection, mapping supervised by anatomical contours like colon folds, and restricted temporal deformation. On C3VD and C3VD-v2, it reports strong overall results, and on C3VD-v2 against SplaTAM it lowers rotational relative pose error by 20.3% and depth RMSE by 38.4%, while improving PSNR by 0.78 dB.
In briefNemotron 3.5 Bolt 74B-A9B 24GB: A Compact Vision-Language Model for Local AI Agents
SummaryThis paper introduces Nemotron 3.5 Bolt 74B-A9B 24GB, a hybrid Mamba-Transformer vision-language model designed for speed and capability per GB of memory in local agentic tasks. It is derived from the Nemotron 3.5 Super VL 120B-A12B family and uses Star Elastic compression with 2-bit quantization-aware distillation, along with a released LiLiCorr speculative drafter. The authors report higher accuracy than models with similar memory and speed on benchmarks such as GDPVal v2, BrowseComp, and Terminal-Bench 2.1, reaching 84 TPS on RTX Spark.
Ganesh Ajjanagadde, Chris Alexiuk, Luke Alonso et al.
In briefJAVIS: Real-time talking agents that speak and show faces together
SummaryThe paper proposes JAVIS, an end-to-end streaming framework in which speech and facial output are generated together from one shared audio-visual representation for real-time conversational agents. The authors also release JAVIS-Syn, a 334-hour synthetic dataset of emotionally rich multi-turn and assistant-style dialogues. Experiments reportedly show about 200x lower initial latency than a strong baseline, along with better lip synchronization and facial expression appropriateness.
In briefGenerative Value Classifier: learning value functions from actions via return-conditioned policies
SummaryThis paper proposes a generative value classifier for Q-learning. Unlike standard approaches that directly regress state-action values from scalar or distributional return targets, it asks how likely an action is under each possible return. The generative component is a return-conditioned policy that can share the actor's architecture, so every action supervises the critic. The method is evaluated on offline and online RL tasks and reportedly outperforms leading discriminative methods, including distributional RL.
In briefUpper limits on causal pieces in spiking neural networks
SummaryThe input space of single-spike neural networks can be divided into causal pieces, and their number gives a lower bound on approximation error for some spiking networks. This paper derives two complementary upper bounds on the number of causal pieces in feedforward single-spike networks. One is a combinatorial bound based on poly-Bernoulli numbers that applies to a broad class of neuron models. The other is a hyperplane-arrangement bound for specific neuron models, with explicit orders given for several of them.
Dominik Dold, Dennis Elbr\"achter, Philipp Petersen
In briefStar Light, Star Bright: Can Multimodal Models Read an Astronomy Chart Quiz?
SummaryThis paper tests how well large multimodal models understand scientific figures using an astronomy visual question answering dataset. The dataset has 2001 partly synthetic sky images, including contour plots, real NASA SkyView telescope images, and synthetic images, with 47357 questions, plus versions with simulated historical page aging. Testing three low-cost proprietary models, the authors found they often fail to answer domain-specific questions accurately. A comparison with 104 verified historical journal figures showed similar median errors in angular extent. Code and dataset are publicly available on GitHub and HuggingFace.
In briefA Unified Theory of Image Scaling: Camera-Optics-Based Scaling Operators and Scale-Spaces
SummaryThe paper derives a system of image-scaling operators, determined by the camera's point spread function, that reproduces optical zoom. The operators compose and invert, covering both upscaling and downscaling. The authors argue that common scaling practice is neither optically faithful nor internally consistent, and that Gaussian scale-space is an idealized special case. The predictions are validated with real captures and noisy optics simulations, and the authors report improved stability for super-resolution and classification.
In briefHas GPT-6-Astra Reached Embodied Intelligence?
SummaryThis report evaluates GPT-6-Astra's robotic manipulation and navigation abilities across 12 benchmarks, including tabletop manipulation, humanoids, mobile platforms, and dexterous hands. It also studies an agent setup pairing a high-level reasoner with a low-level controller, and real-world tests on 8 robot configurations. The authors find that demonstrations for in-context learning are costly to collect and that the model has largely reached embodied competence but not yet practical embodied intelligence.
SenseNova, Joint Research Team, Zhongang Cai et al.
In briefLIBERO-PeRM: A Benchmark for Personalized Robot Manipulation
SummaryThis paper introduces LIBERO-PeRM, a controlled benchmark for evaluating personalized robotic manipulation that separates task completion from preference satisfaction. It includes five task suites (312 tasks) with paired demonstrations, with and without preferences, and automatic preference-satisfaction metrics. Experiments show that current VLA policies struggle with zero-shot personalization, while supervised fine-tuning (SFT) makes preferences learnable, but preference generalization, decomposition, and conflict handling remain difficult.
In briefGen4P: Video Generators Can Also Perceive 4D Geometry and Motion
SummaryThis paper presents Gen4P, a unified framework that recovers 4D geometry and motion from monocular video. It repurposes a pretrained video generator into a feed-forward perceiver that predicts requested attributes via a lightweight cross-attention decoder, without iterative denoising. A frozen video-depth LoRA is added as a geometry prior to improve performance. The authors report state-of-the-art or competitive results on 3D point tracking, video depth, and camera pose estimation.
In briefWhich Tokens to Merge? Diffusion-Dynamics Token Merging for Faster Image Generation
SummaryThe paper proposes DaTMe, a training-free method that speeds up diffusion image generation by merging redundant tokens. It decides which tokens can be merged using a dynamics-aware importance map that combines inter-step changes in the predicted clean image with local high-frequency structural content. Experiments on PixArt-α and FLUX report that the approach preserves visual fidelity across the whole image.
In briefSoftCache: Adaptive Step-Size Caching for Faster Diffusion Transformers
SummaryDiffusion and flow models create images or videos through many iterative steps, each requiring a full pass through a large network, which is costly. SoftCache is a training-free method that replaces the usual yes-or-no decision of whether to recompute with a graded one, using how far a caching signal stays below its threshold to set how far each cached step advances. Steps can skip several scheduled points when there is ample headroom and shrink back as a refresh approaches. On FLUX for images and HunyuanVideo and Wan2.1 for video, it achieves speedups from 3.5x to 6.9x and outperforms all baselines on four fidelity metrics.
Shimon Malnick, Nadav Orenstein, Ohad Fried et al.
In briefT2-Z: Test-time optimization of environment latents for robot world models
SummaryThis paper proposes T2-Z, which adds a learnable environment latent to a robotic world model to capture unobserved factors such as friction and mass. Training alternates between optimizing the world model and the environment latents, while at test time the model is frozen and only a new latent is optimized. Across 7 simulated settings and 4 real-world tasks, the authors report that object prediction error falls from 15.20 to 7.00 and action success rises from 33.73% to 72.96%. Ablations indicate a single test-time interaction trajectory can suffice to infer the latent.</summary>
In briefPerceive, Predict, and Plan: A Unified Model for Egocentric Hand Manipulation
SummaryThis paper introduces P³-WAM, a single model for first-person (egocentric) dexterous manipulation that handles three modes: perceiving hand motion and camera paths from video, predicting visual futures under given actions, and planning action trajectories with their expected visual outcomes. It builds a structured 3D action representation into a pretrained video diffusion transformer. The authors report that unified training improves all three abilities and outperforms specialized baselines on 3D hand reconstruction and action-conditioned video prediction across four datasets.
In briefCoDA: Co-Evolving Simulation Data Engines and Robot Action Policies
SummaryCoDA is a framework in which a simulation data engine and a robot action policy improve each other through a shared memory of environment construction, policy failures, and learning progress. A vision-language critic's diagnoses are turned into generation and training programs, and Text2potential converts them into staged potentials that give dense supervision for policy updates. Across four manipulation task families, out-of-distribution success rose from 5.3% after supervised fine-tuning to 27.6%, 9.5 points above the strongest closed-loop baseline.
In briefGrounding Generative Verification in 3D Content
SummaryThis paper introduces 3DVerifierBench, a benchmark of 10 tasks that tests whether models can judge 3D content, explain rejections, and respond to changes in specifications and candidates. The authors build a data pipeline that produces 14,126 training examples with contradiction explanations and counterfactuals. They also study how verification feedback can support candidate selection and recursive refinement for 3D generation.
In briefPathFM: Fast generative forecasting of multivariate future paths
SummaryThis paper introduces PathFM, a time series forecasting model that uses flow matching to generate complete future trajectories, capturing how future values depend on each other. Through distillation to one or two sampling steps and shared history attention caching, it generates multiple paths in roughly the time of one quantile-model query. It is competitive on marginal accuracy on fev-bench and TIME, and the authors' new Forecast Path Score Bench (fps-bench) reports a higher aggregate path score than the evaluated external baselines.
In briefPACE: A Pareto-based method to compress visual tokens in native multimodal models
SummaryNative multimodal LLMs tokenize image pixel patches directly, which creates heavy visual token redundancy that these models tolerate under compression. PACE estimates token importance by decomposing attention along height and width, and uses Pareto optimization to balance performance, efficiency, and spatial coverage when choosing compression depth and retention ratio. Experiments report near-lossless performance at 25%-50% visual token retention, with some benchmarks improved and over 2x FLOPs reduction.
In briefDK-VLM: Using Geometry as Position to Improve Spatial Reasoning in Vision-Language Models
SummaryThis paper introduces DK-VLM, a vision-language model designed to better understand 3D spatial relations across multiple views. It uses Dual-Kernel Attention, combining the pretrained semantic attention with a geometric attention that is driven by estimated 3D points and camera poses, fused through a learned gate. The authors also propose Reference-Frame Composition supervision to help the model update its reference frame across successive spatial judgments. The reported results include the highest average score across six spatial benchmarks for DK-VLM-8B.
In briefOmniVerifier-C1: A decomposed graded verifier for rule-based generative world modeling
SummaryThis paper introduces a generative visual verifier that splits complex prompts or environment rules into fine-grained components and scores each one, replacing simple True/False judgments. It is trained with a decoupled reinforcement learning strategy that jointly optimizes sub-question binary judgments and global alignment scores. The authors report that the graded verifier outperforms binary verifiers on text-to-image tasks and structured world-modeling tasks such as Maze and Sudoku. They also argue that pairwise verification is more accurate than pointwise scoring when ground-truth images are available.
In briefRethinking Routing Exploration in Modern Mixture-of-Experts Models
SummaryMixture-of-experts (MoE) models scale parameters by activating only some experts per token. Modern designs combine fine-grained expert segmentation with shared experts, which the paper argues create an exploration dilemma: shared experts reduce token representation variability and thus routing exploration, while fine-grained segmentation enlarges the routing space and needs more exploration. The proposed MirrorMoE shrinks the exploration space by token-adaptively suppressing unnecessary experts. The authors report consistent performance gains across model scales.
In briefSUPRA-TEX: A Method for 3D Generation with Super-Aligned Textures
SummaryThis paper presents SUPRA-Tex, a method that generates 3D assets from a single image by dressing a given mesh with relightable materials while preserving the reference image's visual details. It separates visible-material estimation, surface alignment, and observation-guided 3D completion, first predicting PBR maps in the reference view and encoding them as latents that an interleaved flow transformer uses as fixed anchors. The authors report that, trained only on public datasets, it achieves state-of-the-art results in geometry-conditioned texturing and end-to-end asset generation.
In briefFDE-Gym: A benchmark for AI delivery to real customers
SummaryThis paper introduces FDE-Gym, a benchmark of 95 end-to-end delivery tasks from 20 customer scenarios across 18 sectors, adapted by industry experts from real projects. Agents interact with scripted stakeholders and operate on production-like systems in simulated customer environments to deliver working solutions. Across 13 model configurations, the best achieved a mean score of 0.56 and resolved 14 of 95 tasks. The main bottleneck identified is discovering unstated constraints.
WhyProposes a new benchmark scoring delivery by customer business uplift across 95 tasks and 13 configurations, identifying hidden-constraint discovery as the main bottleneck.
In briefSpeculative Forcing: Faster Autoregressive Video Generation by Skipping Unneeded Denoising Steps
SummaryFew-step autoregressive video generators run every token through the final denoising step, even when many tokens are already close to their final output. This paper introduces Speculative Forcing, which uses intermediate clean estimates as drafts, predicts which tokens can be accepted early and frozen, and keeps their cached keys and values as context. It evaluates a training-free verifier and a learned verifier. On LongLive and Lingbot-World 2.0, the method preserves full-step quality while reducing denoising steps per token and increasing generation speed.
Hyunwook Choi, Seungryong Kim, Dongyoon Han et al.
In briefMLExpert: A Generalist Agent Model for Autonomous Machine Learning Engineering
SummaryThis paper proposes a post-training method for autonomous, long-horizon machine learning engineering and uses it to train MLExpert, a 35B Mixture-of-Experts agent model. The approach combines the Kaggle-MLE Factory for building executable tasks, hierarchical agentic supervised fine-tuning, and Windowed Interaction RL, which assigns shared outcome-based rewards to multi-turn decisions between experiments. On MLE-Bench Lite, the average medal rate rises from 37.9% for the base model to 72.7%, and the authors report gains on other tasks as well.
In briefGeoState: Learning view-independent 3D scene states for predicting how a driving scene evolves
SummaryThis paper proposes GeoState, a 3D world modeling method that separates how a scene changes over time from how it looks under a particular camera. A scene tokenizer merges features from multiple views into a view-decoupled scene state, which a geometry decoder queries with camera rays. An autoregressive model predicts future scene states, and the authors report better results than prior models on Waymo and zero-shot generalization to KITTI and nuScenes.
SummaryThis paper introduces 3DCodeVerse, an ecosystem that lets coding agents create editable, verifiable 3D content at scale. It includes 3DCodeVerse-2M, a dataset of about two million executable 3D programs, and a multi-agent framework that automatically checks and revises generated code using execution, spatial measurements, and visual feedback. The authors report that fine-tuning on the dataset improves open-source models' 3D coding and that their harness wins 90% of pairwise human comparisons against Claude Design (3D Object).
In briefPlayPolicy: Learning Dexterous Manipulation from Supervised Play
SummaryThis paper introduces PlayPolicy, a framework called Dexterity from Supervised Play (DSP), in which a robot learns a play policy from minutes of human play via behavior cloning and then autonomously collects hours of interaction data to train a world model. The same policy proposes candidate actions for model-predictive control, and a closed-loop goal-conditioned policy is distilled through self-play in the learned model to speed up inference. In real-world experiments on pushing and in-hand pose reorientation with a multifingered hand, the method outperforms hindsight behavior cloning by 69.6% in success rate and matches reinforcement learning using 490 times less interaction data.
In briefGraph-based lemma retrieval for Lean theorem proving
SummaryThis paper introduces LeanGraphSearch, a retrieval method for Lean 4 that builds a graph from Mathlib's forward dependencies and traverses multi-hop dependency chains to find supporting lemmas. It propagates relevance from semantically matched declarations to their dependencies and reranks the expanded candidates. The authors report up to 12% better recall than LeanSearch v2 on MathlibQR and improved pass rates in an iterative proving loop on three theorem proving benchmarks.
In briefRobust reinforcement learning for adaptive cancer treatment under uncertainty
SummaryThis paper proposes a robust reinforcement learning framework for learning adaptive treatment policies that generalize across patients and stay reliable despite biomarker measurement errors and shifts in disease dynamics. It models each patient as a hidden context within robust contextual Markov decision processes. In a simulator calibrated to clinical data for metastatic prostate cancer, the learned policies substantially outperform an existing adaptive therapy method. The authors suggest robust RL is a promising principled approach for learning reliable adaptive treatment policies.
In briefEvaluation Awareness Beyond English: Whether LLMs Behave Differently When Tested in Indic Languages
SummaryThis paper studies whether large language models can tell they are being tested, a behavior called evaluation awareness, and checks whether it also appears in five Indic languages alongside English. Four instruction-tuned models from three families were tested with cues saying they are monitored or under a safety test. When told they were being tested, models refused harmful requests more often, in all models and languages. In Indic languages, refusals of harmful requests rose by 14.4 percentage points on average, matching English, while refusals of benign Indic requests rose from 6.9% to 11.3%.
WhyFirst systematic study of evaluation awareness in Indic languages, across four models and six languages, showing safety tests may misrepresent real-user behavior.
In briefKnowGPO: Knowledge-Grouped Policy Optimization for Search Agents
SummaryThis paper addresses how deep search agents, which make many sequential search decisions, can be fairly credited for their performance. It proposes KnowGPO, a reinforcement learning method that breaks each query into atomic knowledge conditions, groups trajectories by which conditions they satisfy, and rewards the knowledge gained at each step. The authors report gains of 3.8 to 4.2 points over outcome-only and other process-supervision baselines on Xbench, BrowseComp-ZH, and GAIA, with an 8B model comparable to most 32B open-source systems.
In briefTransformer design for continuous diffusion language models
SummaryThis paper notes that most continuous diffusion language models reuse transformer architectures built for causal LLMs or image generation, which were not designed for them. The authors systematically study transformer architectures for these models and propose LangDiT, which combines feature and query-key normalization, wider noise conditioning, likelihood-shortcut gradient routing and calibration, and confidence-gated training self-conditioning. Used as a plug-and-play backbone, LangDiT lowers generation perplexity across model formulations in unconditional 1024-token generation on OpenWebText.
In briefACEdit: Stable multi-turn image editing via asymmetric conditioning
SummaryImage editing models perform well on single edits, but their quality degrades over repeated editing turns as small distribution shifts accumulate. The paper proposes a training scheme in which the editor learns from its own previous outputs, supported by a distillation objective with a teacher editor that provides reliable supervision. The teacher stays anchored to the original clean image and receives a compound editing instruction. Experiments report substantially reduced degradation over long multi-turn sequences while preserving editing quality and instruction adherence.
In briefTwice Upon a Time: Complementary Visual Views for Time-Series Reasoning
SummaryThe paper proposes giving time-series language models two complementary visual views: line plots that keep absolute measurements and coordinate frames, and normalized delay embeddings that remove amplitude rescaling and offsets while showing temporal structure. A dual-view model built on Qwen3-VL-8B raises TSRBench accuracy from 44.5% with line plots alone to 53.2%. It also reports 89.9% on TSExam and 86.8% numerical accuracy on multivariate ChatTS alignment.
Hedi Zisling, Liran Nochumsohn, Ilan Naiman et al.
In briefDense4D: a model that tracks every body surface point across video frames
SummaryDense4D is a feed-forward model that establishes dense 4D correspondence for the body, clothing and hair from monocular RGB images, without a body template. It predicts Human Point Maps and depth for each image, which let points hidden in a reference frame be located from other frames. Experiments on real people report higher tracking accuracy and closer agreement with scanned surfaces than the evaluated baselines.
In briefTraining models to generate diverse, high-quality scientific ideas from prior papers
SummaryThis paper studies generating research ideas from source papers with large language models, aiming for ideas that are each high quality and clearly distinct from one another. The authors train with a set-level reinforcement learning objective that weights each idea's quality by its distance from earlier ideas in the set. The paper reports that its trained 8B model scores higher on creative utility than two compared models on a contamination-free benchmark, with proof-of-concept success comparable to one of them.
Manya Wadhwa, Pranav Narayanan Venkit, Hiroaki Hayashi et al.
In briefLearning self-orchestration in decentralized agent swarms via market-based training
SummaryThis paper introduces market-based post-training (MPT), which jointly learns action policies and self-orchestration for decentralized agent swarms. Agents bid to express willingness to participate, and the bids determine who acts and how collective outcomes assign local training credit. MPT outperforms monolithic post-training and alternative coordination strategies on embodied crafting, robot navigation, and deep research tasks. Trained swarms also adapt faster to related tasks, scale to larger populations, and degrade less under agent dropout.
WhyProposes market-based joint learning of actions and self-orchestration in decentralized swarms, with comparisons and ablations across three tasks; influence is mainly within multi-agent learning.
In briefAVC-Reward: Event-Centric Checking of Whether Generated Audio and Video Match
SummaryNewer models can generate complex audio-video scenes, but existing metrics were built for simpler tasks. The paper proposes an evaluation framework that analyzes audio and visual events separately, cross-checks them across modalities, and combines the results into a three-level consistency score. It implements this as AVC-Reward, a trainable reward model, and builds AVC-Benchmark to measure agreement with human judgment. The authors report that AVC-Reward aligns best with human judgment among automatic methods and improves a joint generation model when used as a reinforcement-learning signal.
In briefNot Just a Subject: Adaptive Scene-Specific Variance for Visual Storytelling
SummaryDiffusion models can generate coherent image sequences from narrative prompts, but existing methods focus mainly on keeping identity consistent and adapt poorly to varied scenes. The paper proposes EOS-Story, a training-free framework that balances identity consistency with scene-specific adaptation in both text and vision spaces. It also extends the Consistory+ benchmark with diverse framing scenarios, and experiments show competitive or superior performance on both benchmarks.
In briefFour Gemini 4 Argon AI agents autonomously beat Minecraft together
SummaryResearchers placed four Gemini 4 Argon agents in a sandboxed Minecraft survival world with the goal of defeating the Ender Dragon. The agents controlled players through the open-source Mineflayer library and coordinated via in-game chat and shared memory, planning and writing all their own programs without assigned roles or human-written game logic. Starting from a random world on Easy difficulty, the team wrote 2,783 programs and defeated the dragon in 27 hours and 14 minutes, dying 605 times.
In briefLearning to Steer LLM Agents and Multi-Agent Systems
SummaryThis paper addresses goal drift, where LLM-based agents and multi-agent systems stray from their overall objective during long, tool-using tasks. The authors propose a steering agent that monitors execution and intervenes when drift occurs, using a fast low-cost path to decide when to act and a slow high-cost path to decide how. Both paths are trained jointly with reinforcement learning. Experiments on information-seeking search tasks show improved performance and efficiency for single-agent and multi-agent systems.
In briefMindEval2: Clinically grounded patient simulation and rubrics for testing AI in mental health support
SummaryThe paper introduces MindEval2, an evaluation framework with two parts: Mental Health Quality Rubrics, developed with PhD-level psychologists, and MindSim, an LLM-based patient simulator with profiles, tools, and a memory component for multi-turn interactions. The authors evaluated 18 proprietary and open-weight models. They found that frontier models have improved on surface-level quality and boundaries, but guidance calibration remains the main failure mode across all models.
John Mendon\c{c}a, Daniel Tschernutter, Jos\'e Pombal et al.
In briefGrid-Bench: Can LLMs Predict How Discrete Worlds Change?
SummaryThe paper introduces Grid-Bench, an executable benchmark where eight families of spatial operations transform discrete worlds represented as 2D, 3D, and 4D grids. Models must predict the complete final grid from an initial grid, explicit rules, and a textual event history. Evaluating six LLMs, the authors find that complete-state accuracy drops sharply as histories grow, and that models with similar exact-match accuracy can differ substantially in the number of incorrect cells.
In briefVelox: Spatial Memory for Efficient Video World Models
SummaryVelox is a training-free attention router that selects region-level context for video world models using camera frustum slabs built from streaming geometry. Historical blocks are represented as depth-interval frustum slabs, and well-covered regions have their attention importance softly adjusted by slab overlap. At chunk 48, it speeds up attention by 4.36 times and denoising by 2.43 times. On the full WBench navigation split, it stays within 0.15 average points of the dense LingBot-Fast-Cam baseline.
In briefARC-SORT: Adaptive, Robust, Compensated Tracking of Multiple Objects in Sports Video
SummaryThis paper presents ARC-SORT, a training-free, real-time tracker for following multiple players in sports broadcasts. It uses the uncertainty of an Interacting Multiple Model Kalman filter to adapt each track's search region and corrects at most one identity swap when two tracks cross. The authors report HOTA results on SportsMOT and DanceTrack test sets and run at about 50 frames per second on an RTX 3060 GPU.
In briefKnowledge Linearly Readable Before Fine-Tuning Generalizes Better After It
SummaryThis paper distinguishes "active" facts, which support generalization beyond their own recall, from "inert" facts, which can be recalled but do not support such generalization. The authors build new relations by relabeling the answers of existing ones (such as place to continent), fine-tune models on them, and test whether the new labels transfer to held-out subjects sharing the original answer. Across eleven language models and 70 relations, how linearly readable the original relation is in pretrained representations strongly predicts this generalization, with within-model Pearson correlations of 0.87 to 0.92.
WhyConstructed relabeled relations show linear readability of original relations strongly predicts fine-tuning generalization (r≈0.87–0.92), robustly controlled across eleven models and 70 relations.
In briefTesting whether LLM agents can really plan, using formal semantics
SummaryThis paper examines claims that LLM-based agents can plan, monitor, and correct their actions, which are usually supported only by end-to-end success rates or LLM judges. The authors ground natural-language tasks in hidden PDDL models that formally check action applicability, goal attainment, and plan existence. Across five domains, four models, six prompting strategies, and three agent architectures, the best setup solved 78% of solvable tasks offline but 40% online, and recognised unsolvable tasks in under 10% of episodes.
WhyTests LLM planning against PDDL semantics, revealing large online-offline and unsolvable-task gaps across domains, models, and strategies.
Oriol Miro-Lopez-Feliu, Oscar Mir\'o L\'opez-Feliu, Victor Gimenez-Abalos et al.
In briefAI Consilium: Benchmark for LLM Clinical Multidisciplinary Second Opinions
SummaryThis paper introduces a multidimensional benchmark for testing how well large language models give clinical second opinions, judged on factual grounding, coherence, actionability, and safety. It uses 300 diagnostically resolved NEJM–MedQA cases and evaluates eight single-model pipelines that produce independent and interactive consilium reports. A physician-calibrated LLM judge scored 3,040 final reports, finding that model identity mattered more than response format. Structured multidisciplinary synthesis improved scores over pre-consilium primary-role opinions, especially in problem prioritization, factual grounding, and reasoning coherence.
Alexey A. Bokov, Ivan I. Panin, Nikolay B. Nechaev et al.
In briefKeeping Old and New Audio Watermarks Compatible
SummaryThis paper studies backward compatibility in zero-bit neural audio watermarking, where embedders and detectors are updated separately over time. The authors propose training a new detector alongside a chain of frozen legacy embedders, and introduce an Angular Diversity Loss (ADL) that pushes each new embedder into a distinct output region. The authors report that the combined approach restores near-perfect true negative rate without measurable loss in detection accuracy or audio quality.
Admir Krni\'c, Kosta Pavlovi\'c, Petar Nedi\'c et al.
In briefDescribing and Recognizing Activities of Every Person in Videos, Given Their Locations
SummaryThis paper introduces GHADAR, a task that describes the characteristics and activities of every person in a video, given each person's location. The authors build a new open-set captioned dataset from AVA-Actions and propose constraining cross-attention masks in vision-language models during training. Their method outperforms state-of-the-art VLMs, and the paper also proposes a VLM-based evaluation that compares video content directly with predictions.
Jay N. Paranjape, Jurijs Nazarovs, Rohan Sarkar et al.
In briefMulti-Target Few-Shot Prompting with Diffusion Language Models
SummaryThis paper proposes multi-target few-shot prompting with diffusion language models, where examples guide infilling of several masked regions on a shared canvas. This suits batch edits that depend on both left and right context. The authors also use selective masking to derive a likelihood-ratio reward that favors example-supported infills. On few-shot benchmarks and software refactoring case studies, demonstrations improved infilling accuracy, with further gains in several settings.
In briefTowards Continual Learning: Reducing Forgetting via Subspace Tuning
SummaryFine-tuning large language models to learn new skills can erase existing abilities, a problem known as catastrophic forgetting. The paper proposes Subspace Tuning (SST), which limits each weight update to the bottom singular subspace so principal directions are left unchanged. The authors report that protecting up to 99% of the weight subspace improves retention while keeping strong target-task performance, and that SST yields 2–4× less KL drift from the base model than baseline recipes.
In briefLanguage models learning nominalized action constructions across 96 synthetic languages
SummaryThis paper examines how different grammatical configurations affect language models' learning of action nominal constructions (ANCs) and related structures, and compares these learning preferences with typological patterns in natural languages. The authors build synthetic corpora for 96 languages using a grammar-engineered method and train GPT-2-small models from scratch. The models are tested with minimal pairs covering 34 grammatical phenomena. Results show parallels with natural languages in basic clause configurations but divergences in complement systems and ANC strategies.
In briefCAT-alyzing Honesty: Continuous adversarial training to make LLM honesty transfer across tasks
SummaryThe paper addresses large language models that still give dishonest outputs, such as claims contradicting their knowledge or unsupported by the given context, even after alignment training. It introduces CATCH, a continuous adversarial training method that uses a few hundred targets built from self-reporting scenarios to push models toward honest answers. The method is evaluated on Llama-3.3-70B-Instruct and Qwen3-3.6-27B across the MASK, Among Us and AuditBench benchmarks, and is reported to outperform prompting and supervised fine-tuning while preserving utility better than honesty activation steering.
Yuan-Chen Chang, Niklas Herbster, Gauthier Gidel et al.
In briefEliteReplay: Keeping Diverse Solutions in RL Post-Training via Quality-Diversity Search
SummaryThis paper proposes EliteReplay, a post-training method that pairs GRPO with an elite archive built through quality-diversity search. Solutions from the frozen base model are collected before training, and a replay loss on this archive is combined with GRPO during training. On several text-to-SQL tasks, it achieves higher pass@128 than all baselines and retains more of the base model's correct strategies. On the ambiguous-question benchmark AmbiQT, it covers both interpretations nearly twice as often as the strongest baseline.
In briefImproving General Agents: How Scaffolding Still Matters
SummaryThis paper evaluates a Codex-based general-purpose agent on ARC-AGI-3, AGI Maze, and Terminal-Bench. Changing one instruction to encourage reporting difficulties and returning control improves mean scores on ARC-AGI-3 and AGI Maze, with mixed small changes on Terminal-Bench. The general agent still trails a specialized ARC-AGI-3 reference, and moving reset decisions to a scripted controller recovers part of that gap. The authors also note the gains come with higher token cost.
In briefLearning rational strategies for temporal goals in multi-agent systems
SummaryThis paper studies rational synthesis in stochastic multi-agent systems with unknown dynamics, where each agent has a probabilistic temporal objective written in PLTL_f. Solution concepts such as Nash equilibrium are imposed as constraints on the joint strategy profile. The authors give conditional convergence guarantees for tabular Nash-Q learning and validate the approach on 8 general-sum concurrent stochastic games.
Omar Adalat, Edwin Hamel-De le Court, Francesco Belardinelli
In briefAsking Smarter Questions: Targeted Active Preference Learning for Underspecified Rewards
SummaryThis paper addresses reward learning in cases where human teaching covers only some task features well, and robots can detect which features were poorly supervised. Rather than asking humans to re-teach, the authors propose an informed prior that directs preference queries at under-resolved features and explanations that focus the human's attention on those features. In simulation, the informed prior outperforms standard active preference learning by up to 30.8% when few queries are allowed. A user study found that both the prior and the explanations significantly improve reward recovery and amplify each other.
In briefRouting Policy Optimization (RPO): a reinforcement learning method that adds exploration through attention routing in image and video generators, then restores the original deterministic sampler.
SummaryThis paper proposes Routing Policy Optimization (RPO), which turns attention routing in visual generators into a stochastic exploration action at selected denoising levels. Two rollouts that differ only in their sampled attention routes give a reward difference used as the learning signal for the routed transitions. At inference, RPO restores dense attention and keeps the original deterministic sampler. The authors report improved generation quality on text-to-image and text-to-video tasks and nearly 4x higher training efficiency than Flow-GRPO-style methods.
In briefLEASH: Uncertainty-Aware Escalation and Oversight for Computer-Use Agents
SummaryThis paper proposes LEASH, a training-free framework in which a weaker open-weight computer-use agent acts by default and hands control to a stronger model only when its latent uncertainty signals this is needed. After handoff, a shadow observer keeps oversight of the strong black-box agent. On OSWorld, LEASH improves over standalone weak agents and beats random escalation at matched escalation rates. With Qwen3.5-35B-A3B it reaches 73.66% task success, close to always-on Claude Opus 5 at 72.90%, while cutting inference cost by 19.2%.
In briefBenign recursive self-distillation can cause gradual misalignment
SummaryThis paper studies the safety of recursively training a model on its own outputs via self-distillation. A safe model, given a harmless cue, generates answers that serve as the teacher for the next round. Experiments on Qwen3 and Phi-4 using SORRY-Bench and StrongREJECT show that misalignment on harmful queries grows with each round, a effect the authors call Progressive Misalignment. The authors also find benign prompts that trigger this failure without harmful training data.
WhyReveals benign recursive self-distillation causing progressive misalignment, with gradient-found benign prompts; evidence spans two model families and two benchmarks.
Karim Abdel Sadek, Federico Di Gennaro, Fanny Yang
In briefAttributing Harness Evolution Gains to Overfitting and Test-Time Scaling
SummaryThe paper introduces Harness-Delta Attribution (HDA), an evaluation protocol that splits gains from automatically optimized prompts, tools, and orchestration code into overfitting, test-time scaling, and a residual term. Applied to 11 executor models and three frameworks across math, coding, creativity, and agentic tasks, it finds that in 21 settings with significant search-set gains, overfitting and test-time scaling explain 54% and 14% of the gain on average. The authors report that search-set gains often fail to transfer to held-out tasks and that validation signals do not prevent overfitting, motivating attribution-aware evaluation.
In briefAuditable Multi-Agent Motion Planning Under Legal Rules Using Partially Ordered Games
SummaryThe paper proposes a framework for multi-agent motion planning under legal requirements, based on games with partially ordered preferences (posetal games). Scenario-independent legal guidelines rank abstract legal concerns, from which agent preferences and a legal game are automatically derived for each concrete scenario. Tested on maritime navigation across seven encounters involving up to four vessels, the framework certifies rule violations as either forced by the legal hierarchy or unsatisfiable. The authors argue it offers an auditable route from partially specified laws to multi-agent motion planning.
In briefMJev-Spatial: A Fast Single-Pass Model for Spatial Understanding via Attention Matching
SummaryThe paper proposes MJev-Spatial, a multimodal model that handles choice, numerical, and point-based spatial tasks as one classification problem. It scores candidates with the backbone's existing query and key projections, adds no parameters, and outputs each answer in a single forward pass without generating text. Built on Qwen3.5-4B and trained on 12.3M converted spatial question-answer pairs, it reports the highest overall score (65.64%) among eight tested models and runs faster than its autoregressive base model.
In briefLibrarian Agents: training LLMs to organize long documents in advance for future questions
SummaryThis paper proposes Librarian Agents, where a builder agent converts raw documents into a hierarchical filesystem of synthetic files and reference links for a downstream answering agent. The workspace construction is formulated as a Markov decision process and the builder is trained with reinforcement learning to maximize downstream answering accuracy. Reported results show a 33 percentage point gain over the strongest retrieval and precomputation baselines on BrowseComp-Plus, and zero-shot generalization to unseen domains on BRIGHT with a 31-point average accuracy improvement.
In briefUnsupervised clustering of city buses using shape and color features
SummaryThis paper proposes an unsupervised method that groups public transport vehicles by fusing body-shape features from a frozen DINOv2 Vision Transformer with handcrafted HSV color histograms, followed by UMAP projection and K-Means clustering. It was evaluated on 7,972 bus images extracted from about 1,460 hours of video on an arterial road in Mexico City. With a 65% morphology and 35% color fusion, the pipeline produced 14 clusters separating distinct routes and vehicle types and improved the precision of the target cluster over morphological features alone.
In briefX-RIC: An agentic RAN intelligent controller benchmark for improving reinforcement learning
SummaryThis paper proposes X-RIC, an explainable framework that integrates agentic AI with reinforcement learning in O-RAN's RAN Intelligent Controller. It combines automated behavioral cloning, a transformer-based Telecom Language Model for reasoning and interpretability, and closed-loop co-adaptation in the Near-RT RIC. The authors also introduce BARX, an evaluation framework for latency, compliance, and determinism, and validate X-RIC on SLA-aware network slicing in simulation and an O-RAN-compliant testbed. They report 89% operational cost reduction while maintaining SLA compliance.
In briefBeyond Pairwise Repulsion: Batch-Level Diversity in Diffusion Sampling via Determinant Maximization
SummaryStrong guidance in text-to-image diffusion models tends to make outputs look alike, which hurts synthetic data augmentation that needs varied samples. The paper proposes DetNO, which optimizes initial noise to maximize the log-determinant of an RBF Gram matrix over one-step predictions, encouraging diversity across the whole batch. It uses an anchor-based fidelity term and a single trade-off hyperparameter, requiring no training or sampler changes. Experiments on SD1.5, SDXL and SD3 show leading intra-batch diversity and statistically significant few-shot classification accuracy gains.
In briefSCOPE: Semantic Bounded Credit Optimization for Prompt Engineering in Text-to-Image Generation
SummaryThis paper proposes SCOPE, a reinforcement learning method that trains a language model to rewrite prompts for a frozen text-to-image generator. It groups tokens into semantic spans such as objects, counts, attributes, and relations, and pools policy log-probability shifts within each span into a shared update weight, without needing separate span rewards or extra image renders. Compared with vanilla GRPO, SCOPE reports average relative gains of 9.48% on GeneVal and 7.62% on T2I-CompBench across Qwen-Image, SD3, and FLUX.1-dev. Ablations indicate that both the semantic boundaries and the within-span pooling contribute to these gains.
In briefAuthor-involved AI agent audit of reproducibility for a machine learning conference
SummaryThe paper proposes a collaborative reproducibility audit in which authors' declarations, recorded in a structured checklist, serve as explicit inputs to an AI agent's investigation. A two-stage GitHub workflow links agent checks, author clarification and revision, and agent reassessment, with human reviewers approving requests. Deployed across 36 conference papers and 133 tables and figures, the system raised 20 issues after human review, and further checks supported closing all of them, though full reproduction remained unverified in some cases.
In briefBaselines and score centering in policy gradient estimation via covariance
SummaryThe paper studies off-policy reinforcement learning for post-training large language models, where importance sampling ratios are regularized by truncation, masking, or clipping. It treats the covariance between reward and importance-weighted score as the estimation target, which restores invariance to constant reward shifts and matches the policy gradient when the ratio is unregularized. It analyzes the roles of the group-mean baseline and score centering at both sequence and token levels. It concludes that with sequence-level importance sampling score centering adds nothing once the group-mean baseline is used, while with token-level regularized ratios the group-mean baseline alone generally introduces bias.
In briefOnline Political Narratives in Moldova's 2024 Presidential Election
SummaryThis paper studies how political narratives spread on X, Facebook, and Reddit during Moldova's 2024 presidential election. It combines multilingual stance detection, coordination analysis, and causal inference to examine positions on European integration, Russia, and the main candidates. The authors report positive, statistically significant average treatment effects of peer influence across major stance categories, suggesting that amplification and peer exposure help reinforce online political narratives.
In briefTopology-inspired fine-tuning to improve wind-field super-resolution
SummaryScientific super-resolution models are usually trained with pointwise losses, which can miss extrema and connected structures that matter for interpretation. The authors compare pretrained CNN and GAN wind-field models, finding that pointwise and merge-tree metrics favor the CNN while Wasserstein persistence-diagram distances favor the GAN. They fine-tune the CNN with differentiable topology-inspired losses and fixed persistence-pair supervision. On a 168-field benchmark, the final objective lowers mean persistence-diagram distances by 25–31% and gains 1.30 dB PSNR, while merge-tree distance improves on only 91 of the 150 fields where persistence distances improve.
In briefFPG-Seg: Label-free retinal vessel topology segmentation
SummaryRetinal vessel segmentation usually needs extensive manual annotation by clinical experts, which is very time-consuming. This paper presents FPG-Seg, a pipeline combining classical vesselness filtering, Bayesian deep learning and topology-aware graph reconstruction to segment retinal vessel topology without manual labels. On four widely used benchmark datasets it reaches a clDice of 0.74 to 0.78, and it provides the first performance benchmark of an annotation-free vessel segmentation method on FIVES.
In briefDenoising Textures: Why It Needs Its Own Approach
SummaryThis paper studies how state-of-the-art image denoisers handle texture images, where fine patterns often get over-smoothed. The authors find that performance gains on textures come only from fine-tuning on texture datasets, and that this fine-tuning causes the models to catastrophically forget the original dataset. They argue that texture-aware denoisers are needed for faithful texture reconstruction and synthesis.
Saptarshi Mandal, Oindrila Haldar, Chirayata Bhattacharya et al.
In briefBitSparse: Lossless compression of sparse activations to cut LLM training memory
SummaryThe paper presents BitSparse, a GPU-resident lossless format that compresses sparse ReLU and ReLU² activations stored during LLM training and restores them exactly for the backward pass. It supports BF16 and FP8 activations, with optional sign-bit packing for higher compression. The authors report per-token VRAM reductions of 31% on Modded-NanoGPT and 30% on Nemotron-H, with overheads of no more than 1% at practical sequence lengths.
In briefExplore once, evaluate many: evaluating multiple decision policies in linear bandit problems
SummaryThis paper studies how to estimate the value of K target policies at the same time in stochastic contextual bandits with linear rewards. The authors propose PRISMA, an algorithm that achieves accuracy ε with probability at least 1-δ using reward-free policy queries. It also gives an upper bound on the number of context draws and reward observations required.
In brief"Not X, but When": Diagnosing a Chinese LLM Writing Habit by Contextual Appropriateness
SummaryThis paper examines the Mandarin construction "not X, but Y" (bushi X, ershi Y), a recurring expression in text written by language models. The authors introduce a decision tree that classifies each use as substantive, optional, or defective, and compare five commercial models with a human reference. They find that models fail more often on appropriateness than on correctness, and that some reward models and an LLM judge favor the construction.
In briefRouting Weak Model Answers to Stronger Ones Using Model Disagreement
SummarySmall language models on edge devices may need stronger models for hard cases, but deciding when to escalate usually requires ground-truth correctness signals that are costly to obtain. This paper proposes routing based on disagreement between weak and strong model responses, which can be measured without ground-truth supervision. Across agentic and single-turn tasks, the approach balances routing cost and accuracy, and it can outperform methods trained to predict correctness even when such signals are available.
In briefSurvICL: A Foundation Model for In-Context Survival Analysis
SummaryThe paper finds that existing tabular foundation models do not beat the decades-old Cox model in survival analysis, which predicts event timing from censored data. It introduces SurvICL, which predicts survival curves in-context, is pretrained only on synthetic data, and uses a plain quantile regression loss. On 63 datasets held out from design decisions, SurvICL wins 82.9% of pairwise integrated Brier score comparisons against 71.3% for the runner-up, and is the first model to rank above Cox on small datasets.
WhyProposes a synthetic-prior-pretrained in-context survival model beating 20 baselines on a large benchmark; impact mainly within survival analysis.
Julie Alberge, Linus Bleistein, Judith Ab\'ecassis et al.
In briefMMSearch-OPD: A training method for long-horizon multimodal search agents
SummaryThis paper introduces MMSearch-OPD, a method for training multimodal deep search agents that gather evidence over long investigations. It uses a Refine pipeline to expand reasoning chains from existing simple image-question pairs, and combines reinforcement learning with on-policy distillation. Experiments report gains from the data refinement and further improvements from joint training after SFT.
In briefQuery-efficient rejection sampling to correct parallel decoding bias in masked diffusion models
SummaryMasked diffusion models speed up generation by unmasking several tokens at once, but sampling them independently ignores their correlations and biases the output. The paper proposes the Single-Increment Rejection Sampler (SIRS), which estimates the log acceptance probability from one random increment per query and uses Rao-Blackwellization to reduce clipping. The authors bound the KL divergence between the data distribution and SIRS output by a clipping term plus the model's training error.
In briefTraining-free method merges specialist models into one using only forward passes
SummaryThe paper proposes a training-free way to combine several specialist models, each tuned from the same base model, into a single student model. Each linear layer is solved as a least-squares problem using second-moment statistics from a few hundred rollouts, requiring only forward passes. The authors report that filtering out answers truncated in repetition loops substantially affects results. On Olmo-3-7B, the fused generalist retained most of each specialist's gains across seven mathematics and code benchmarks.
In briefOS-Helix: A Computer-Using Agent That Combines GUI Actions and Code Execution
SummaryThis paper presents OS-Helix, a multimodal computer-using agent trained end to end to combine GUI actions and code execution in one action space. The authors build executable Ubuntu environments by pairing desktop applications with file artifacts and generate verifiable tasks, then apply SFT followed by online reinforcement learning. They propose Step-balanced GSPO, which weights optimization by action step rather than token count. On OSWorld-Verified, OS-Helix-27B-RL reaches a 66.44% success rate.
In briefWaveLLM: LLM Entropy Modeling in the Wavelet Domain for Lossless Image Compression
SummaryThe paper proposes WaveLLM, which decomposes images into multi-level wavelet subbands so that the prediction problem becomes more structured and statistically homogeneous, then uses a LoRA-adapted Llama-3.2-3B model to estimate probabilities autoregressively. It reports 2.51, 2.24, and 2.27 bpsp on Kodak, CLIC Pro, and DIV2K, respectively. These results outperform a PixelCNN-based codec using the same wavelet transform and recent LLM-based lossless compressors on all three benchmarks.
In briefDesigning for Hardware Diversity: Meta's End-to-End Co-Design for AI Training on AMD GPUs
SummaryThis paper describes how Meta optimized recommendation model training on AMD Instinct MI350X GPUs, using NVIDIA B200 as a comparison. It proposes three reusable principles: hardware-model co-design, an analytical performance model for warmup-free kernel selection, and extended agentic workflows for porting from CUDA. Compared with a directly ported baseline, these methods improve end-to-end training throughput by up to 1.5x.
In briefHow Much Physics Vision Can Learn from Audio: Distilling Damping Structure
SummaryThe paper compares frozen audio encoders with a single impact video frame encoded by CLIP for ordering objects by their quality factor Q, a measure of damping. Audio embeddings reach a Spearman correlation of 0.73 versus 0.28 for the frame, and this advantage largely holds on withheld materials. Distilling the audio teacher into CLIP through a projection head recovers a useful part of the gap, and under leave-one-material-out evaluation a four-frame input raises correlation from 0.234 to 0.338.
In briefSkill-decoupled on-policy expert distillation for merging vision-language model abilities
SummaryThis paper proposes SCOPED, a method for combining separately trained capabilities in vision-language models. Capability-specific experts are trained, and explicit skills are extracted through reward-guided reflection. During on-policy distillation the skills guide expert supervision but are unavailable to the unified student at inference. SCOPED outperforms the backbone and mixed-capability RL across eight multimodal benchmarks.
In briefFutureLift: Refining Motion Predictions with Bidirectional Future Graphs
SummaryFutureLift is a backbone-agnostic refiner that corrects multimodal trajectory proposals for autonomous driving motion forecasting. It represents each proposal as a heterogeneous graph linked to agent histories, map elements, alternative modes, and other agents' futures, and uses bidirectional temporal edges to revise predictions in one pass. It improves frozen proposals on Argoverse2 and the Waymo Open Motion Dataset and ranks first in minADE and minFDE among compared standard-input methods on the WOMD leaderboard.
In briefFaster, better layout splitting for chip patterns combining multi-patterning and self-assembly
SummaryThis paper studies how to split chip layouts for a lithography process that combines multiple patterning with directed self-assembly (DSA-MP). The authors add a new conflict graph model that includes edge-to-edge spacing rules for DSA templates, which earlier methods overlooked. They also use a graph neural network to locally refine the decomposition. Experiments show better solution quality and faster convergence than prior methods, with robustness across layout scales and technology nodes.
In briefGeneralization guarantees for randomized block coordinate descent on nonconvex learning problems
SummaryThis paper studies how well randomized block coordinate descent (RBCD), a common large-scale optimization method, generalizes when the objective is nonconvex. Prior theory mainly addressed optimization convergence. The authors derive high-probability bounds on gradients of empirical and population risks and show how tuning the iteration count balances optimization and generalization. They also give dimension-independent bounds for a specific loss structure.
In briefTRACE: A Training Objective That Restores Broken Connections in Thin-Structure Segmentation
SummaryThin, elongated structures such as paths or vessels can be broken by a few local segmentation errors. TRACE is a training objective that uses a target skeleton to define required connections and, for each one, finds the easiest valid reconnection under the current prediction, supervising its bottleneck with the matching skeleton point. Experiments on six 2D and 3D benchmarks show improved connectivity and topology, with strong segmentation accuracy and low training overhead.
In briefBeyond Downstream QA: Rethinking How Agent Memory Is Evaluated
SummaryCurrent agent memory benchmarks use downstream question-answering accuracy as a proxy for memory quality, but this conflates memory correctness with retrieval and generation quality. The authors introduce FEB-30, a benchmark of 30 fact-evolution trajectories across 6 diagnostic categories and 3 difficulty levels, with documents generated from known trajectories to give deterministic ground truth without human annotation. Applied to a representative memory system, intrinsic and extrinsic evaluation diverge by an average of 16 percentage points, and the system correctly classifies memory operations only 20% of the time.
Sebastian-Antonio Toma, Jean de Kernier, Mike Ply et al.
In briefDecentralized LLM workflows run through sequential auctions among self-interested agents
SummaryThe paper proposes a scalable framework in which self-interested agents in a decentralized market solve multi-step tasks through sequential auctions. It shows that both perfect competition, where agents are exact clones, and no competition at all lead to market failure. The authors evaluate four bidding algorithms across three auction protocols to study how heterogeneous agents behave under imperfect competition. They report that imperfect competition sustains agent profit margins while keeping task success rates high at low cost.
In briefOpenGCUA: A fully open technology stack for general-purpose computer-use agents
SummaryThis paper presents OpenGCUA, an open project that aims to release the full stack for generalist computer-use agents (GCUAs) that complete tasks through visual and programmatic interfaces. The authors add code execution and harness-to-model decision transfer, which lift the mean partial score by 19.6 points and cut steps per task by 53.0% on OSWorld-v2 across eight models. They release their implementation, prompts, evaluation trajectories, and preliminary real-device harness and cross-device SDK.
In briefSafe Multi-Robot Pursuit with Smarter Communication Timing
SummaryThis paper addresses safe coordination among resource-constrained robots that cannot share their states continuously. Each agent computes the reachable set of its neighbors from the last received state and uses control barrier functions to choose a collision-avoiding control input. In pursuit-evasion simulations, preemptive communication reduced messages by 77.5% compared with periodic and event-triggered baselines, with zero collisions.
