News
· 2026-09-30
Hugging Face has introduced FailBank, a four-stage self-evolving framework that converts runtime feedback into persistent policy improvements for Vision-Language-Action (VLA) models. FailBank employs a fixed safety module to generate counterfactual corrections and leverages LoRA updates to optimize model performance. Evaluated on the VLA-Arena benchmark, FailBank improved task success rates by 6.9 to 8.5 percentage points and significantly reduced policy-induced cumulative costs, outperforming e
News
· 2026-09-30
Hugging Face has released Dream4ACT, a novel world model for joint video-action modeling across multiple embodiments. By introducing a shared visual action interface called 'action views,' which leverages URDF-based forward kinematics, Dream4ACT unifies action representations across different embodiments while preserving their specific articulated geometries. The model supports forward dynamics, inverse dynamics, and joint observation-action generation within a single jointly trained framework.
News
· 2026-09-30
Hugging Face has introduced Diptych, an AI-assisted system designed to enhance reference listening in music production. Diptych enables users to define comparison scopes across entire tracks or selected segments and provides structured audio features alongside scope-specific AI interpretations. Evaluated with 12 musicians and expert listeners, Diptych demonstrated strong usability and the ability to surface additional differences with expert support, highlighting AI's potential in creative and d
News
· 2026-09-30
Hugging Face has introduced JevSpawn, an innovative agentic inference framework designed to address the efficiency bottlenecks of existing large language models (LLMs) in complex task execution. JevSpawn connects natural language task specifications with finite probabilistic exploration, enabling parallel action spawning, feedback-driven branch selection, representation revision, and recovery from retained alternatives. Its key advantage lies in reducing repeated generation and context computati
News
· 2026-09-30
Architect-Ant is an innovative framework for automatically generating furniture layouts in architectural floor plans. It learns professional furnishing patterns using a coordinate-based DSL for editable layouts and optimizes them by integrating geometric and functional constraints. The AntPlan dataset, containing 505 real professional floor plans with dense annotations across 92 object classes and ten residential room categories, provides high-quality data for model training. Experiments demonst
News
· 2026-09-30
EgoTools is the first comprehensive suite for understanding tool use in real-world egocentric videos, consisting of two main components: EgoTools-Data, a large-scale corpus of 100 hours of tool-centric recordings with synchronized audio, dense captions, reasoning-heavy narrations, and supplementary 3D information; and EgoTools-Bench, a diagnostic benchmark with 1,000 QA pairs across four tracks covering perception, geometry, procedure, and causal reasoning. Experimental results show that current
News
· 2026-09-30
Hugging Face has introduced the SEEK framework, designed to address memory challenges in robotic manipulation under partial observability. SEEK combines three complementary memory mechanisms to retain historical evidence and track execution states. Evaluations on the HIDE benchmark reveal significant limitations in existing policies, while SEEK's memory augmentation substantially improves task success rates in both simulation and real-world experiments. This framework offers a more reliable solu
News
· 2026-09-30
AIGeeksGroup introduces DexPolicy, a novel approach to optimizing reinforcement learning (RL) in dexterous manipulation by explicitly controlling the scale of exploration. DexPolicy anneals the exploration scale from broad to narrow as a function of training steps while keeping loss, architecture, reward, and optimizer settings fixed. Experiments demonstrate significant improvements in target success rates across various scenarios, such as increasing the success rate from 25.0% to 85.0% on a Rea
News
· 2026-09-30
BeyondSCe is a novel zero-shot robotic grasping system designed for event-referential settings, where robots identify and localize target objects for grasping based on event history and the current scene. The system excels in handling occluded targets by combining event priors with scene geometry to select optimal camera viewpoints. In real-robot experiments, BeyondSCe achieved grasp success rates of 76% and 77% for initially visible and occluded targets, respectively, outperforming the stronges
News
· 2026-09-30
GraphForge is an innovative evidence-graph based framework designed to address the challenges of data synthesis and task verification in agent training. By assembling workspaces of real files and building evidence graphs over their relationships, GraphForge ensures that task requirements are grounded in verifiable data. Initial experiments show that fine-tuning Qwen3.6-27B on trajectories generated by GraphForge leads to significant performance improvements across benchmarks like OpenHands, Work
News
· 2026-09-30
Fireworks AI has launched a new reinforcement learning (RL) trainer and rollout engine that co-optimizes training and inference processes. This ensures alignment across numerics, kernels, and mixture of experts (MoE) routing, enabling teams to scale RL applications more effectively and achieve higher performance and reliability in complex tasks.
News
· 2026-09-30
DAGent is a Directed Acyclic Graph (DAG)-based multi-agent framework that introduces Evaluate-then-Grow incremental planning to optimize task execution. By incrementally expanding the task graph based on confidence and uncertainty signals from completed nodes, DAGent enhances the efficiency and accuracy of complex task execution. It outperforms existing open-source baselines across multiple benchmarks and demonstrates strong adaptability and scalability across different scales.
News
· 2026-09-30
ThyVoice has introduced the Speaker Identified cpWER (SI-cpWER) metric to evaluate persistent speaker attribution in meeting transcripts. This metric, which scores a corpus under a single global speaker-ID assignment, addresses the limitations of existing metrics that fail to measure the consistency of speaker identities across meetings. Tests on the CHiME-8 NOTSOFAR and CHiME-6 datasets demonstrate that ThyVoice outperforms all evaluated commercial cascades in terms of speaker attribution accur
News
· 2026-09-30
MemLife is an innovative multimodal memory system designed to address the critical challenges in long-term egocentric video memory processing. It constructs entity-grounded, first-person text episodes and retrieves them via a time-indexed agentic reader, outperforming the strongest training-free baseline by 4.6% to 12.0% across four long-horizon benchmarks. To further enhance memory quality, the researchers introduced MemOpt, a reinforcement learning framework that optimizes the memory writer to
News
· 2026-09-30
Hugging Face's research team has introduced SimpleB2T, a novel approach to enhance non-invasive brain-to-text decoding performance. By preventing the neural network from exploiting timing intervals as shortcuts, SimpleB2T significantly improves the accuracy of decoding text from brain signals. Experiments demonstrate that SimpleB2T outperforms traditional methods on both synthetic and real brain data, making existing strategies much more effective. This breakthrough opens new possibilities for t
News
· 2026-09-30
Hugging Face's research team introduces EvolvingNav, a novel approach to address the challenges of persistent navigation for embodied agents in evolving environments. By constructing a time-indexed belief from timestamped 3D object histories and integrating an event-driven filter with RGB-D evidence updates, EvolvingNav significantly improves navigation success and search efficiency. The method demonstrates strong performance in both simulation and real-robot experiments, particularly under lear
News
· 2026-09-30
Hugging Face's research team investigates the safety implications of latent communication in multi-agent systems. The study reveals that even benign link training can increase harmful compliance relative to text-based communication, posing a threat to system safety. Attackers can amplify this effect by optimizing links on harmful query-response pairs or poisoning otherwise benign training data. A reinforcement-learning-based attack is developed, which rewards harmful compliance alongside benign
News
· 2026-09-30
Hugging Face has introduced DyRAD, a novel method for radar novel view synthesis in dynamic driving scenes. DyRAD models dynamic scenes using static background reflectors and motion-tracked dynamic point reflectors to render complete range-azimuth-Doppler (RAD) tensors. Unlike traditional methods, DyRAD leverages radar's Doppler effect as both a rendered output and supervision for object tracks and employs a fixed analytic point-spread function (PSF) to prevent sensor-induced spread from being b
News
· 2026-09-30
Self-evolving search agents build their own training curricula by jointly optimizing a question generator (proposer) and a solver. However, this closed loop can introduce co-cheating, where the proposer and solver increasingly agree on shared errors, improving internal rewards without a corresponding gain in external correctness. A new method, CrossFit, is proposed to address this issue. It partitions the proposer's source documents into two groups and uses an auxiliary solver trained only on th
News
· 2026-09-30
NVIDIA's research team introduces PivotOPD, a novel framework designed to address pivotal mistakes in multi-turn agent interactions. PivotOPD jointly trains agents to prevent pivotal mistakes and recover from the resulting states, significantly improving task success rates. In experiments across ALFWorld, WebShop, and search-based QA tasks, PivotOPD demonstrated superior performance with Qwen3-1.7B and Qwen3-8B models, outperforming the strongest baseline by 5.5% on ALFWorld. Additionally, it bo
News
· 2026-09-30
Mid-Harness is a novel approach that enhances the reliability of terminal agents by allocating test-time compute resources at the boundary between the model and the harness. By sampling and verifying candidate actions before execution, Mid-Harness reduces the risk of environmental disruption caused by poor commands. Experiments on TerminalBench-Lite using a TMAX-9B generator and verifier show that Mid-Harness improves Pass@1 from 50.00% to 68.03% with 8 sampled actions. Furthermore, combining ac
News
· 2026-09-30
Hugging Face has released PatchHolmes, an innovative two-phase patch retrieval system designed to improve vulnerability management workflows. PatchHolmes leverages a hybrid retriever and an agentic inspection loop that reads the top-100 candidates listwise, outperforming existing methods like Favia and IRCoT by significant margins in GitHubAD benchmarks. The system operates on a frozen open-weight model without requiring fine-tuning or external APIs, highlighting its potential for efficient and
News
· 2026-09-30
Anthropic has introduced SkillSeek, an open-source tool designed to address the bottleneck in agent skill retrieval. SkillSeek employs a standard information retrieval (IR) approach with a two-stage retrieval mechanism (a BGE-based bi-encoder and a small cross-encoder) to achieve efficient skill selection. In the SkillsBench benchmark, SkillSeek achieved parity with existing LLM-mediated retrieval loops at virtually no extra cost, reducing the per-task cost from $51.30 to $27.54. This demonstrat
News
· 2026-09-30
Hugging Face has released the Agent Error Dataset (AED), a comprehensive dataset comprising 50,228 error-diagnosis pairs across 33 environments, 19 harness families, and 23 policy models. AED aims to facilitate failure analysis and error-aware post-training for agent systems by providing detailed error diagnosis data. The five-stage Agentic Error-to-Training (AET) pipeline automates failure collection, diagnosis generation, and correction validation. Experiments demonstrate that fine-tuning with
News
· 2026-09-30
Hugging Face has released EviRover, a novel perception agent designed to address the limitations of traditional visual perception models in handling fine-grained visual details or knowledge-intensive tasks. EviRover employs an interactive perception process to gather information and is trained using supervised fine-tuning followed by agentic reinforcement learning. Its performance surpasses the backbone model by an average of 30 points on the EviLens benchmark and demonstrates strong results acr
News
· 2026-09-30
ROBOCOACH is a world-model-guided coaching framework for robot skill learning, addressing the inefficiency of skill composition in long-horizon robot manipulation tasks. The framework employs a Route-Imagine-Diagnose-Improve (RIDI) loop to simulate failure scenarios using a shared action-conditioned world model (COACHWORLD) and guide expert updates and demonstration requests. Experiments across two simulation suites and two real-robot platforms show that ROBOCOACH significantly improves task suc
News
· 2026-09-30
RSIGame is an autonomous agentic game development framework with recursive self-improvement, designed to address the challenges of reliably improving generated games beyond a playable version. The framework organizes development into complementary local and global loops, enabling evidence-grounded revision and long-term quality tracking. Across 140 GameCraft-Bench tasks, two game engines, and five generators, RSIGame consistently improves game quality under matched development budgets. Notably,
News
· 2026-09-30
arXiv introduces the Memory Adaptation for Task-Conditioned Execution (MATE) framework, a novel approach to enhance memory adaptation in embodied agents. Traditional memory systems focus on construction and retrieval but often fail to address semantic relevance and historical success, leading to issues when retrieved experiences contain incompatible actions or inappropriate structures. MATE addresses these challenges by removing obsolete control contexts, extracting condition-action-effect trans
News
· 2026-09-30
A new research paper on arXiv introduces 'Environment Steering,' a novel approach to improving the safety and reliability of LLM-based agents. This technique enforces safety policies during runtime by modeling the agent and execution state as database tables, tracking data flows, and checking them against declarative policies. When violations are detected, the system provides policy- and context-specific feedback to steer the agent toward safe trajectories. Experiments on AgentDyn demonstrate th
News
· 2026-09-30
Hugging Face has introduced the Game-Guided Skill Discovery (GGSD) framework, which leverages self-play in games to discover motor skills directly playable by humans. The framework employs a hierarchical agent that competes against its past selves, with a high-level policy selecting from a discrete skill set and a skill-conditioned low-level policy learning the corresponding behaviors. After training, users can control the agent through the same discrete skills, enabling the composition of compl