News
· 2026-09-29
Video-RSI is a novel framework designed to enhance the performance of video understanding agents through recursive self-improvement. The framework leverages the agent's own language model to revise its execution harness, revisiting training videos to test competing failure explanations with additional observations, grounding proposed changes in evidence beyond the existing trace. Experimental results demonstrate that the evolved agent improves accuracy while processing fewer frames and achieves
News
· 2026-09-29
Hugging Face's research team conducted a study on the direct transfer of a co-evolved communication protocol from a 2D simulation to a 3D physical environment without retraining network weights. The study found that the transfer was partial and asymmetric, with one robot succeeding in reaching the food source in only one out of thirty trials, while the other failed in all. The results suggest that successful transfer of emergent communication depends not only on preserving the signaling process
News
· 2026-09-29
Hugging Face has introduced the SimpleICL framework, addressing fundamental challenges in robotic in-context learning (ICL) by clearly defining learning objectives and resolving prompt ambiguity. This minimalist and reproducible framework, equipped with a visual prompt encoder and a low-cost data collection protocol, achieves strong performance in both simulated and real-world environments without requiring massive pre-training or specialized data infrastructure. The team has also open-sourced t
News
· 2026-09-29
Hugging Face has introduced PAMI (Part-Anchored Motion for Interaction), a novel framework for generating text-conditioned full-body human-object interactions. PAMI leverages body-part anchors to localize object motion and employs a coarse-to-fine hierarchical generation approach, significantly improving the fidelity and accuracy of human-object interactions. Experiments on the InterAct dataset demonstrate that PAMI achieves a 14.5% higher contact recall rate compared to previous state-of-the-ar
News
· 2026-09-29
Hugging Face's research team introduces ProgressCompass, an innovative approach to address context-dependent progress estimation challenges. By integrating existing Progress Reward Models (PRMs) with general-purpose Vision-Language Models (VLMs), ProgressCompass provides the necessary context to agents, significantly improving their progress estimation accuracy in complex tasks. Experimental results show a 63% reduction in progress error and a 76% increase in rank agreement, offering a new techn
News
· 2026-09-29
Hugging Face has introduced Video2Skill, a benchmark designed to evaluate the ability of Vision-Language Models (VLMs) to discover and reuse skills from video streams. The benchmark covers robot tabletop manipulation and human kitchen activities, testing three core capabilities: temporal event localization, event grouping based on transformations, and decision-making on skill reuse. The study reveals that many VLMs struggle with event grouping, and model scaling does not consistently improve per
News
· 2026-09-29
Hugging Face has released Proactivity-Gym, a simulation-based evaluation testbed for designing, developing, and evaluating proactive Large Language Model (LLM) agents. The framework is grounded in three core principles (3T): Task Capability, Temporal Allocation, and Trust, and organizes its design space around five dimensions: task scope, anticipation horizon, activation trigger, processing timing, and intervention depth. Proactivity-Gym evaluates proactive assistance across interactions using m
News
· 2026-09-29
Hugging Face has introduced the WEFT (Whole-system Evolution For Tool-use Post-training) framework, which aims to enhance model performance in complex tasks by building scalable agentic interaction systems, execution-driven self-evolution, and stable post-training. WEFT scales across environment breadth, task complexity, and interaction diversity, iteratively using execution traces and state evidence to revise components and evaluate changes through fresh rollouts. Extensive experiments demonstr
News
· 2026-09-29
Hugging Face has introduced Modal Kinetic Typography, a novel technique that animates vector glyphs to convey semantic concepts while maintaining legibility. By leveraging the natural vibration modes of the glyphs and employing a frozen video diffusion model to supervise the amplitudes and phases of these modes, the method produces smooth and coherent animations. This approach addresses the limitations of prior work, such as outline tearing and jitter, by ensuring smooth motion along the outline
News
· 2026-09-29
Hugging Face has released HelixWorld, the first real-time interactive audio-visual world model that natively co-evolves visual scenes and camera-grounded spatial stereo sound under user interaction. Built on a high-fidelity spatial audio-visual dataset and pre-trained with a bidirectional teacher conditioned on 6-DoF camera trajectories and user actions, HelixWorld achieves low-latency causal interaction through online trajectory distillation. It sustains drift-free joint audio-visual rollouts a
News
· 2026-09-29
Hugging Face has introduced MotorMind, a novel robot manipulation framework that enables general-purpose vision-language models (VLMs) to perform effective zero-shot robot manipulation without relying on external models or tools. MotorMind connects VLM-proposed mid-level actions to deterministic robot control and feedback, with asynchronous monitoring and background memory updates. It achieves a 66.7% success rate on the LIBERO-PRO benchmark and 95% average success on a real xArm6 robot across d
News
· 2026-09-29
OpenAI has introduced Dots, a proactive assistant designed to help users manage complex projects and daily tasks efficiently. Dots offers the ability to work across tasks proactively, providing support and maintaining user control as work progresses. This design emphasizes enhancing task efficiency and collaboration through AI while ensuring the user remains in command. The release highlights OpenAI's ongoing innovation in the field of intelligent agents, paving the way for new advancements in A
News
· 2026-09-29
Hugging Face has introduced RLE-Bench, a novel benchmark designed to evaluate the broader engineering capabilities of robot learning agents. The benchmark spans four key robotics development workflows: interactive control, policy learning, perception and estimation, and mechanical design. It employs diverse task-specific metrics to assess agent performance and aggregates them into an overall RLE Index, enabling systematic comparisons across multiple capability dimensions. Beyond performance rank
News
· 2026-09-29
Hugging Face has introduced OTRetarget, a novel approach for joint retargeting of human and multi-object motion to humanoid robots. This method leverages entropic optimal transport to transfer surface interactions across human, robot, and object geometries while preserving contact consistency. Evaluations on the OMOMO dataset show that OTRetarget achieves a robot-object interaction Jaccard score of 87% and a depth error of 8.7 mm, outperforming the baseline OmniRetarget. The approach has been su
News
· 2026-09-29
Hugging Face has introduced Rules to Tools (R2T), a novel tool designed to enhance the efficiency of AI agents in scientific programming tasks by providing executable checks for scientific requirements. Experimental results demonstrate that agents using R2T significantly outperform those relying solely on textual descriptions, completing 29 out of 30 tasks compared to 26 for the control group. Notably, R2T excels in specific tasks such as task 17, 77, and 11, where it successfully flags and repa
News
· 2026-09-29
The Hugging Face team introduces the Retrieval-Augmented Skill Optimization (RASO) framework, which leverages an external skill corpus as prior knowledge to optimize agent skills. RASO adapts relevant knowledge from existing skills to the target task and harness through Cross-Harness Adaptation, addressing mismatches in both domain and constraints. It consists of two stages: Retrieval-Augmented Skill Initialization (RASI) and Retrieval-Augmented Skill Update (RASU), enabling efficient skill cons
News
· 2026-09-29
MILO (Meta-evolutionary Island Orchestration) is a novel framework for agent harness design that co-evolves agent execution control structures and the strategies used to discover them. This approach significantly enhances the performance and efficiency of complex tasks by leveraging hierarchical lineage memory, per-island mutator agents, and adaptive search orchestration mechanisms. MILO outperforms eight state-of-the-art harnesses and six search methods across benchmarks like Terminal-Bench 2.1
News
· 2026-09-29
LangChain has announced the integration of Jev, an AI agent, into its LangSmith evaluation platform. Jev is designed to provide faster, more cost-effective structured feedback by evaluating agent behavior traces across production runs, datasets, and regression tests. This release enhances the efficiency of AI agent evaluation and reduces operational costs, marking a significant advancement in LangChain's suite of tools for AI development and assessment.
News
· 2026-09-29
The research team at EPFL has introduced TERRA, an end-to-end pipeline for terrain-aware retargeting and control of musculoskeletal locomotion. TERRA recovers task-relevant support geometry from kinematic trajectories by combining terrain priors, estimated contacts, and negative free-space evidence, while considering anatomical, tendon-continuity, and contact constraints during retargeting. Experiments demonstrate that TERRA significantly improves terrain accuracy, reduces anatomical and interac
News
· 2026-09-29
SkillGym introduces an automated pipeline for training skill-use agents by constructing verifiable environments, collecting trajectories, and fine-tuning Large Language Models (LLMs). The system crawls skills from the internet, filters those with reproducible workflows, and uses a builder-reviewer pipeline to create difficulty-controlled tasks with reference solutions and executable verifiers. SkillGym has built 6,800 environments and collected 19,000 verified successful trajectories for supervi
News
· 2026-09-29
Hugging Face has introduced SMART (Self-evolving Multi-Agent system for long-foRm subtitle Translation), a novel multi-agent system for long-form subtitle translation. SMART employs dynamic routing and a Mixture-of-Agents layer, integrating tools for terminology verification, subtitle constraint validation, and contextual retrieval. During test-time training, it builds persistent series-level memory and uses a judge-refiner loop to score candidates and update agent prompts without retraining und
News
· 2026-09-29
EVOKE is a novel post-training method designed to address the poor transferability of large language model (LLM) agents to unseen environments. By introducing goal diversity at fixed states, EVOKE forces agents to adjust their action preferences, thereby implicitly eliciting their internalized world knowledge. This mechanism significantly enhances the agent's generalization capabilities, task performance, and data efficiency in unseen environments. The introduction of EVOKE offers a new perspect
News
· 2026-09-29
PRISM is a novel real-to-sim-to-real framework that addresses the data scarcity problem in humanoid robot loco-manipulation skill learning by generating a large, diverse set of counterfactual human-object interaction videos from a handful of real videos. The framework reconstructs human and object motions into physically plausible trajectories, enabling the training of a policy that generalizes to unseen scenarios. Experiments demonstrate that the PRISM framework successfully enables robots to p
News
· 2026-09-29
Hugging Face introduces LibraryDesignBench, a two-phase benchmark designed to evaluate how well AI agents can design libraries for other agents. The benchmark assesses the correctness and simplicity of programs written by user agents based on agent-designed libraries across 242 expert-validated programming problems in 15 tasks and four languages. While agents successfully reproduce human-written abstractions in 11 out of 15 tasks, downstream agents often underuse these libraries, opting to reimp
News
· 2026-09-29
Hugging Face's research team introduces Q&D (Questioner and Drafter), a novel framework designed to enhance AI agents' proactivity and information-seeking capabilities. Q&D trains agents to prioritize questions that retrieve more of the required evidence, enabling two forms of proactivity: horizontal (pursuing unstated information identifiable in the current context) and vertical (pursuing needs revealed by earlier evidence). Experiments on multi-hop question-answering benchmarks demonstrate tha
News
· 2026-09-29
Hugging Face's research team introduces EmoRES-TTS, a novel method for emotional speech generation. This approach decomposes emotion vectors into shared and residual components, enabling efficient control of emotional expression without retraining the backbone model. Experimental results demonstrate that EmoRES-TTS significantly outperforms existing methods across various emotion metrics, with improvements of 20.1% in emotion hit rate and 118.8% in rank correlation. Human evaluations further con
News
· 2026-09-29
Hugging Face has introduced CorpusMap, a novel navigation layer for intelligent document retrieval that organizes information around recurring entities within a document corpus. CorpusMap creates an Entity Page for each entity, aggregating relevant information and linking to all documents that reference it, forming a graph between entities and documents. This approach enables AI agents to more efficiently traverse the document collection, discover cross-document evidence, and improve answer qual
News
· 2026-09-29
Hugging Face introduces EpiCon, a shared multimodal memory framework for collective agent learning. EpiCon leverages two independently trained 2B models—a memory controller and a tree self-organizer—to enable experience sharing and accumulation without updating host model parameters. It features mechanisms for refining textual guidance and visual evidence across attempts and selectively integrating visual memory. Benchmarks across four multimodal task domains demonstrate that EpiCon significantl
News
· 2026-09-29
Hugging Face has released EngiWorld, the first benchmark platform tailored for professional engineering autonomous agents. EngiWorld spans six engineering domains (CAD, CAE, CAM, BIM, EDA, and 3D visualization) and 26 professional software platforms, featuring 1,301 expert-curated tasks that test agents' capabilities in handling complex engineering workflows across software and design stages. Its novel artifact-centric evaluation methodology, built on a unified domain-verifier suite, programmati
News
· 2026-09-29
Hugging Face has released HybridCUA, a novel agent framework that combines Graphical User Interface (GUI) and Command Line Interface (CLI) interactions to enhance task execution efficiency for Computer Use Agents (CUAs). The framework employs a data construction pipeline to generate hybrid trajectories and a two-stage training approach involving supervised fine-tuning followed by reinforcement learning with CLI-aware rewards. Experimental results demonstrate that HybridCUA achieves a 53.6% accur