News
· 2026-10-06
Hugging Face introduces MimicX, a policy-in-the-loop framework that refines video-driven humanoid motion tracking using execution feedback. By identifying difficult transitions and affected body regions, MimicX jointly adapts tracking objectives and the reset curriculum for policy continuation. Across four core video tasks, MimicX demonstrates a 25.7% reduction in body-tracking error and a 255.6% increase in execution horizon compared to the fixed reference baseline. Additionally, MimicX-HLoop a
News
· 2026-10-06
Hugging Face has introduced the EMHO (Embodied Agent Harness Optimization) framework, an innovative approach for optimizing embodied agents' planning and interaction capabilities. EMHO works by keeping the underlying model frozen and iteratively refining the agent's control logic (harness) through the analysis of execution trajectories and harness history. This framework enhances how agents monitor progress, utilize vision tools, ground observations, and respond to failures. Additionally, EMHO-M
News
· 2026-10-06
Hugging Face's research team introduces Stateless Language Agents (SLAs), a novel framework designed to address inefficiencies in automated research systems, such as history replay, duplicated work, and resource wastage. By decoupling research state from agents, SLAs ensure that each agent invocation receives a fresh, role-specific context, thereby avoiding redundant computations and enhancing research efficiency. In tasks like software engineering, kernel optimization, and algorithm design, SLA
News
· 2026-10-06
Hugging Face has introduced Hebero, a novel GPU-parallel benchmark for heterogeneous multi-task reinforcement learning (RL) based on Isaac Lab. Hebero enables efficient joint training and evaluation of a single policy across 40 diverse tasks. Scaling experiments demonstrate that increasing parallel replicas per task enhances success rates under fixed time constraints. The framework incorporates Demonstration-Guided Policy Optimization (DGPO), which leverages adaptive behavior cloning (ABC) and i
News
· 2026-10-06
Hugging Face has introduced Station, an AI system designed to tackle the challenges of open-ended scientific discovery. Station incorporates a Supervisor mechanism and periodic Meta Reflection to encourage persistent exploration even in the absence of clear metrics. Experiments show that Station rediscovered 62.7% of the criteria from three recent ICLR papers on average, outperforming existing systems like Codex Multiagent-v2 and AI Scientist-v2. Furthermore, Station demonstrated the ability to
News
· 2026-10-06
Hugging Face has released the CARE (Certifying Acceleration for Vision-Language-Action Inference) framework to address the challenge of maintaining reliability while accelerating inference in vision-language-action (VLA) models. CARE employs a paired rollout approach to evaluate the impact of acceleration strategies on task success rates and provides finite-sample guarantees that the risk of acceleration-induced failures remains below a user-specified budget. In tests on OpenVLA-OFT and LIBERO b
News
· 2026-10-06
Hugging Face has released SPW-Nav, a streaming panoramic world model that generates high-quality 2K 360-degree panoramic videos in real-time based on movement instructions and supports on-the-fly instruction switching. By leveraging spherical rotation decoupling, pose-aligned conditioning, and multi-term memory with a few-step generator, SPW-Nav achieves superior camera-following accuracy and video quality compared to existing panoramic generators. This innovation holds significant potential for
News
· 2026-10-06
Hugging Face's research team introduces Self-Retrospection Distillation (SRD), a novel method that leverages post-hoc experiences to enhance an agent's predictive capabilities before taking action. By distilling insights from completed trajectories, SRD guides the agent's foresight, improving performance in complex tasks. The method demonstrates significant gains across 10 tool-integrated reasoning and long-horizon agentic tasks, particularly in scenarios with scarce reward contrast. In experime
News
· 2026-10-06
Hugging Face has introduced the Agent Plasticity framework, a novel approach to evaluating the efficiency of AI agents in self-improvement through experience. The framework measures how effectively agents convert past experiences into future performance gains, addressing key questions about their ability to learn and adapt in complex environments. It introduces three core evaluation dimensions: the generalization of future performance, the efficiency of acquiring new capabilities, and the identi
News
· 2026-10-06
Hugging Face has introduced WorldSonus, a real-time spatial audio generation framework designed to address the core challenges of sound synthesis in virtual world models. Utilizing a streaming causal autoregressive diffusion architecture, WorldSonus achieves low-latency audio synthesis while supporting dynamic sound event manipulation through an audio-centric captioning pipeline and high-quality stereo supervision. Experiments demonstrate that WorldSonus excels in both acoustic quality and spati
News
· 2026-10-06
The research team at Hugging Face has introduced Gan Jiang, a self-learning agent for powder X-ray diffraction analysis. Built on tools like XMatcher, XQueryer, XDecomposer, and WPEM, Gan Jiang converts analytical experience into executable skills and improves performance by diagnosing failures, revising skill instructions and code, and validating revisions—all without retraining the language model or altering the underlying physical models. Gan Jiang excels in DeltaXRDbench benchmarks, outperfo
News
· 2026-10-06
Hugging Face has introduced CADFather, an autonomous agentic system designed to reconstruct editable CAD models from 3D shapes. The system coordinates complementary tools, using a vision-language assistant to inspect target and intermediate reconstructions, decide on program extensions, tool invocation, proposal generation, and completion timing. CADFather employs learned and algorithmic tools for CAD operation proposals and numerical optimization for parameter refinement. Evaluations on DeepCAD
News
· 2026-10-06
Robo-COP is a novel framework for robotic agents that enables the co-evolution of policies and orchestrators during deployment, enhancing adaptability and task success rates in complex real-world scenarios. The framework curates skill demonstrations from its own executions, fine-tunes policies when recurring failures are detected, and adopts new policies only after verifying their improvements. In simulated tasks, Robo-COP increased the mean held-out success rate from 64.8% to 73.8%, and in real
News
· 2026-10-06
Astra introduces PhysEvo, a novel framework for physical recursive self-improvement (RSI) designed to enhance robotic performance in complex tasks. Built around a single frozen model, PhysEvo employs a task agent and a meta-agent to automate failure diagnosis, skill revision, and correction testing without requiring model weight updates or separate action policy training. The framework develops joint-level control, evidence-driven observation, and reusable manipulation skills. In evaluations acr
News
· 2026-10-06
Hugging Face has released the ProactiveCoach suite, designed to enhance AI assistants' hierarchical understanding and adaptability in task guidance. ProactiveCoach-Instruct provides phased training data, covering task phases, steps, and actions to help AI assistants comprehend task progress and context. ProactiveCoachBench evaluates model performance across different guidance levels and tests their adaptability to changes in guidance granularity. Experiments show that models based on hierarchica
News
· 2026-10-06
Hugging Face has introduced Recursive Game Creator, an intelligent framework designed to enhance the experience of AI-generated games. The framework employs a four-component iterative development process involving Designer, Builder, Player, and Reviewer. The Player component efficiently collects diverse gameplay trajectories through programmatic interfaces, while the Reviewer leverages trajectory-based metrics and visual evidence to evaluate game quality and provide improvement feedback. Experim
News
· 2026-10-06
Hugging Face has unveiled nanoMuse, an open-source personal AI agent framework designed to operate seamlessly across a user's devices. Built on the GPL-3.0 license, nanoMuse enables shared conversations between devices like smartphones and computers, ensuring privacy and security through its Sentinel mechanism. The modular design allows users to select different models, while providing readable memory files and an evaluation suite. This initiative aims to democratize AI agents for multi-device e
News
· 2026-10-06
Hugging Face has introduced Sherpa, a multi-turn reinforcement learning framework designed to enhance the adaptive teaching capabilities of AI teacher models. Sherpa simulates diverse student archetypes with distinct learning preferences and trains teacher models to maximize their learning outcomes. Evaluated on MathTutorBench, Sherpa improved the overall pedagogy score from 52.5% to 79.2% and was preferred over the base model in 79.6% of pairwise comparisons. This innovation paves the way for A
· 2026-10-06
An independent developer, lucidml_lover, has released a new local AI world model on Reddit that can convert images into playable characters in real-time and supports dynamic text prompt switching during runtime. The model, based on a pure Transformer architecture with a novel 'diffusion forcing' training method, has approximately 960M parameters and achieves a peak FPS of 50-60 on an RTX 5090 GPU at low utilization. It is also compatible with other mainstream GPUs and supports real-time control
News
· 2026-10-06
Hugging Face has released VICS-G, a novel reranker designed to address the safety and feasibility challenges in Vision-Language-Action (VLA) policies. By incorporating the distribution of future possibilities, VICS-G significantly reduces the cumulative safety cost of task completion while maintaining high success rates and reducing the number of steps. This approach does not require model retraining or online rollouts, offering an innovative solution for safer task execution.
News
· 2026-10-06
Illuin Tech's research team introduces DAEDALUS, a novel method for bootstrapping reusable agent memory from self-generated tasks without relying on existing tasks or oracle verifiers. DAEDALUS pairs an explorer agent with a solver agent, where the explorer generates challenging yet solvable tasks, and the solver attempts them. By deriving heuristics from solver failures and consolidating accepted heuristics into a memory bank, DAEDALUS improves mean success rates by up to 15.9 points and pass^5
News
· 2026-10-06
Hugging Face introduces NeMo-DCR (Delta-Compressed Refit), an innovative weight synchronization technique for trillion-parameter scale agentic reinforcement learning (RL). By transmitting only parameter changes and ensuring bit-exactness, NeMo-DCR reduces the synchronization time of a traditional 1TB checkpoint from 87.5 minutes to 150 seconds, achieving over 35x efficiency gains. The technology employs fixed affine mappings, residual conversion, and native encoder placement to ensure bit-level
News
· 2026-10-06
Hugging Face's research team introduces SafeActBench, a comprehensive benchmark designed to evaluate the decision-making chains of agents as they move from evidence to action. The benchmark comprises 656 cases across six operational domains and five protocols, ranging from static action judgment to single- and multi-action workflows. The study reveals that agents often fail before execution, either by stopping with incomplete investigation or acting before acquiring the required evidence. While
News
· 2026-10-06
Hugging Face has introduced AdvSim2Real, a novel framework designed to enhance the robustness and task completion capabilities of web agents against adaptive prompt injection attacks. By co-evolving a task curriculum, an injection adversary, and the agent within a frozen web world model, AdvSim2Real enables the transfer of training outcomes from the simulator to real browser environments. In experiments on 150 web tasks, AdvSim2Real improved the agent's task completion rate by 33.6% against unse
News
· 2026-10-06
Hugging Face has introduced Attacca, a novel approach designed to address the challenges of continuous state control in long-horizon task execution for embodied agents. By decoupling goal images from the execution environment and introducing behavioral-phase conditioning, Attacca significantly enhances the performance of agents in complex tasks. Experimental results in Minecraft demonstrate a 1.7-7x performance improvement across various long-horizon tasks, highlighting its groundbreaking potent
News
· 2026-10-06
Hugging Face has introduced MediateRec, a novel benchmark designed to study the decision-making capabilities of personal agents in cross-platform recommendation systems. The system leverages user-authorized cross-platform history to mediate platform recommendations, aiming to balance the trade-off between population evidence and personalized agent decisions. MediateRec incorporates the Personal Attribution Mediation Optimization (PAMO) algorithm, which uses counterfactual masking and platform-re
News
· 2026-10-06
Hugging Face introduces VeriFine, a novel agent harness framework designed to address the verification bottlenecks in self-improving policies. VeriFine enables the co-evolution of policies, training curricula, and judges to diagnose recurring failures, construct adaptive curricula, and optimize policies. The framework incorporates a rubric judge for diagnosing failures and a coactive calibration process for refining judges through human-agent, ensuring continuous improvement in both policy and j
News
· 2026-10-06
Hugging Face has introduced EmbodiedSmith, a novel framework that revolutionizes robotic data generation and simulation training through recursive self-improvement (RSI). The framework unifies asset, scene, and task generation into a cohesive pipeline, enabling autonomous creation and language-driven customization. Its core is an agentic refinement loop where scene generation anticipates task requirements, and task generation guides scene edits, iteratively improving both. This approach enhances
News
· 2026-10-06
Hugging Face has introduced a novel technique called 'Speculative Execution' to optimize the response latency of on-device cascaded voice assistants. By predicting tool requests from partial ASR hypotheses and initiating tool execution during speech recognition, this approach reduces end-to-end response time. Experimental results demonstrate a reduction in median time-to-first-audio from 5.79 seconds to 4.60 seconds and a decrease in standard deviation from 3.49 seconds to 2.81 seconds, making r
News
· 2026-10-06
Hugging Face introduces the ALIVE framework to address the challenge of inserted objects participating in interactions in video editing. ALIVE uses an edited first frame and instructions naming only the added object to enable coherent interactions between inserted objects and the source video content. The framework leverages a curated dataset of 35,800 editing pairs, combining 3D-rendered, model-generated, and real-world videos, and employs a vision-language model (VLM) to predict interaction gu