News
· 2026-10-09
Hugging Face introduces Cortex, a tool designed to test and optimize the rules and skills of coding agents. Cortex validates changes to the agent's harness through replayable tasks, adopting an incremental approach that retains only those modifications that demonstrate a net gain across multiple runs without degrading task performance. Tested on Claude Haiku 4.5, Cortex demonstrated its ability to enhance agent behavior consistency, reduce the presence of useless or harmful rules, and streamline
News
· 2026-10-09
Hugging Face introduces APEX-Harm, a benchmark designed to evaluate the safety of agents operating in shared workspaces. The benchmark simulates complex real-world environments and employs multiple attack categories, such as static/dynamic injections and task-specific script replacements, to assess agent performance in finance, consulting, and legal tasks. Experimental results show that while some models excel in task completion, they still exhibit high attack success rates, highlighting the ong
News
· 2026-10-09
NVIDIA has officially launched Omniverse AI Agents, a suite of developer tools that integrate frontier AI models with NVIDIA Omniverse libraries. This release aims to help developers build simulation applications for tasks such as scenario exploration, failure investigation, and design optimization more efficiently. Developers can direct AI agents through natural-language instructions to perform complex tasks, leveraging Omniverse's GPU-accelerated physics, rendering, and sensor simulation capab
News
· 2026-10-09
Hugging Face introduces Video-Arena, a novel benchmark designed to evaluate the spatiotemporal understanding capabilities of video agents. By transforming competitive games into scalable and verifiable tests, Video-Arena assesses models' abilities to comprehend and track evolving world states from videos. The benchmark includes three task categories: critical event perception, state update interpretation, and resulting state tracking, with evaluations conducted under two protocols: native visual
News
· 2026-10-09
Hugging Face has released WATCH, a novel cross-video pre-training model that learns transferable world-action representations from task-paired human demonstrations across diverse environments. By predicting one demonstration's actions and future frames from another, WATCH encourages the model to learn task-relevant interactions transferable across scenes and demonstrators. Experiments show that WATCH significantly improves the in-context success rate of MimicDroid, particularly on unseen tasks,
News
· 2026-10-09
Hugging Face has introduced a novel Latent World Action Model for GUI agents, shifting the prediction focus from pixel-level reconstruction to latent space representations of action-induced future states. This approach leverages a large multimodal model with dedicated heads for action selection and predictive transition modeling, while introducing an action-aware contrastive objective to mitigate the identity shortcut problem. Extensive experiments demonstrate significant improvements in both su
News
· 2026-10-09
Hugging Face has released Web$hell, a unified interface designed to revolutionize web access for Large Language Model (LLM) agents. By leveraging the WebContract schema, WebCompiler compiles live webpages into actionable objects, while WebCache ensures persistent storage and consistent operations. This approach enhances agents' capabilities in information gathering, browser interactions, and evidence revisiting, all while reducing computational costs. Extensive experiments demonstrate that Web$h
News
· 2026-10-09
Hugging Face introduces Harness^2, a novel framework designed to address the challenges of Recursive Self-Improvement (RSI) for agents in open-world settings. Harness^2 employs a two-tiered recursive mechanism at the task and harnessing levels, enabling policy refinement without relying on ground-truth verifiers and facilitating generalization across tasks. Experiments demonstrate that Harness^2 consistently improves the performance of GLM-5 and Gemini 3.5 Flash across three professional benchma
News
· 2026-10-09
Hugging Face has released BranchOPD (Branching On-Policy Distillation), an innovative approach that enhances agent decision-making and distillation efficiency by introducing branching mechanisms into student-generated trajectories. BranchOPD leverages student uncertainty and teacher-student disagreement to select informative decision points, resamples trajectories from shared prefixes, and applies token-level teacher supervision to both main trajectories and branches. This method significantly i
News
· 2026-10-09
Hugging Face introduces AgentSWE, a novel benchmark for agent software engineering, spanning three critical stages of agent development: creation from natural language requirements, codebase editing, and optimization toward measurable targets. AgentSWE evaluates agents across 25 tasks and reveals that existing agents often fail to meet requirements, highlighting recurring failure modes such as overfitting to visible cases, hollow runtimes, and unsupported claims of completion. This release provi
News
· 2026-10-09
Hugging Face introduces TerminalHorizon, a fully automated data synthesis engine for long-horizon terminal agents. Addressing the limitations of existing environment synthesis methods that often omit critical work, TerminalHorizon reconstructs omitted tasks and progressively expands scenarios to foster exploration and improvement. Fine-tuning Qwen3.5-27B on the resulting dataset demonstrates substantial performance gains across coding-agent tasks, real-world production settings, and general reas
News
· 2026-10-09
Hugging Face has released ThinkDrag, a novel drag-based image editing framework that incorporates visual reasoning for more intelligent image manipulation. ThinkDrag associates user drag instructions with meaningful object- and scene-level transformations and can optionally produce reasoning traces to interpret the intended edits, enhancing both generation quality and interpretability. The framework also introduces Drag Endpoint Markers to ground drag instructions in image space. Experiments dem
News
· 2026-10-09
Hugging Face's new research investigates the emergence of communication between autonomous agents and its relationship with the value of information. The study, conducted using two e-puck robots in a simulated environment, reveals that communication strategies adapt based on the value of information. When lacking guidance, agents tend to adopt referential signaling, while abundant environmental information reduces the need for communication. Additionally, the research indicates that innate behav
News
· 2026-10-09
Hugging Face has introduced Work2Skill, a framework designed to capture human GUI interaction experiences for reusable agent skill induction. The framework includes WorkRecorder, which aligns screen recordings, mouse and keyboard events, and accessibility-tree data to create WorkDemo-350, the first synchronized multimodal dataset for GUI demonstrations. Experiments show that equipping agents with skills induced by Work2Skill significantly improves their performance and success rates in complex G
News
· 2026-10-09
Hugging Face's research team introduces a novel approach to On-Policy Distillation (OPD) based on relative policy differences. This method defines an implicit reward through the log-probability ratio between a proxy expert and a proxy reference policy, enabling efficient knowledge transfer across models of different scales, training stages, or generations. Experiments demonstrate that this approach excels in task-specific training, such as reinforcement learning and supervised fine-tuning, while
News
· 2026-10-09
PKU Value Lab has introduced EmbodiedCLBench, the first benchmark designed to evaluate the continual learning capabilities of self-evolving embodied agents. The benchmark, centered around compositional generalization, assesses whether agents can learn from basic tasks and recombine acquired skills to solve complex, unseen tasks. Experiments reveal that while advanced tasks are hard to solve in a zero-shot setting, experience from basic tasks leads to consistent improvements. However, this improv
News
· 2026-10-09
RSI-Anything is a domain-general framework built on JEPA-Anything that leverages predictive world models to guide actions and learning-data selection, enabling recursive self-improvement. It uses observed feedback to improve policies or workflows and conditionally adapts world models, while component and compatibility checks ensure the retention of updated policy-model pairs. RSI-Anything demonstrates substantial performance gains across diverse domains such as code repair, GUI task completion,
News
· 2026-10-09
ProgramGym is a novel framework designed to address critical challenges in training long-horizon program generation agents. It consists of two components: Program-Env, which constructs verifiable training environments through a feedback-driven loop supporting languages like C, C++, Go, and Rust, and Program-Agent, which integrates an isolated workflow for analysis, implementation, and review with revision management for compilation checks and failure recovery. Experiments demonstrate that Progra
News
· 2026-10-09
LangChain has launched Managed Deep Agents, a novel API and system designed for distributed agents, enabling dynamic management and assignment of emoji responses. This tool enhances the interaction experience between agents and users by providing more flexible and intelligent communication, particularly suited for scenarios requiring complex emotional expression and multimodal responses.
News
· 2026-10-09
LangChain has announced the integration of a model router into its Open SWE toolkit, which optimizes AI model invocation efficiency by intelligently distributing tasks. This innovation reduces the median cost per coding task by 64% while maintaining quality, offering developers an efficient and economical solution for AI-driven programming tasks.
News
· 2026-10-09
LangChain has announced the integration of a model router into its Open SWE toolkit, achieving a 64% reduction in the median cost per coding task without any measurable drop in quality. The model router optimizes AI model invocation efficiency through intelligent task allocation mechanisms, providing developers with a cost-effective and efficient solution. This article also provides guidance on building a similar model router, offering practical insights for AI developers.
News
· 2026-10-09
Postman and AWS have partnered to launch Agent Mode, an AI-native workflow solution built on Amazon Bedrock, designed to enhance developer productivity in API testing, documentation, discovery, and implementation tasks. By addressing challenges such as tool sprawl, context optimization, and leveraging Amazon Bedrock's flexibility and cross-region inference capabilities, Agent Mode provides a scalable and intelligent solution for developers worldwide. This integration demonstrates the potential o
· 2026-10-09
Whistle is an ultra-lightweight speech-to-text model requiring only 16.9MB of storage. It maintains high efficiency while significantly reducing resource requirements, offering a new solution for speech recognition applications on resource-constrained devices. Its compact size and high performance make it highly suitable for IoT devices, mobile applications, and edge computing scenarios.
News
· 2026-10-09
A new study published on arXiv introduces Keyframe Mnemonics, a novel self-supervised method for behavior cloning (BC) in non-Markovian environments. The method identifies a set of information-critical observations (mnemonics) from randomly sampled past observations and uses them as a reward signal for keyframe selection, addressing the limitations of traditional recurrent and attention-based models in capturing long-term dependencies. Experiments demonstrate that mnemonic-conditioned BC policie
News
· 2026-10-09
A new arXiv study investigates the effectiveness of small LLM agents in operating under explicit wall-clock time budgets, focusing on both adherence to allocated runtime and productive use of time. The research identifies critical gaps in time awareness, action duration prediction, and strategy mapping, leading to failures in budget adherence. Proposed solutions include harness-based mechanisms and reinforcement learning with budget-aware rewards, such as GRPO, which achieves near-perfect adhere
News
· 2026-10-09
A new study published on arXiv introduces Plan-and-Patch, a framework that enhances planning and repair capabilities for long-horizon agents. By leveraging a diffusion language model (dLLM), the framework generates structured plans and repairs selected regions while keeping surrounding steps fixed, improving efficiency. In experiments without task-specific training, the diffusion approach achieved a 53.7% success rate in plan repair, nearly doubling the performance of autoregressive models (27.0
News
· 2026-10-09
arXiv has introduced the Trajectory-Grounded Instruction Translator (TGIT), a novel approach to improve the performance of aerial vision-and-language navigation (VLN) agents. TGIT translates user-provided, intent-driven instructions into agent-executable commands by learning from trajectory outcomes. This method significantly boosts the success rate (SR) of weak input commands from 11.33% to 37.93% on the OpenFly navigator and successfully transfers to real human instructions, achieving an SR of
News
· 2026-10-09
arXiv introduces Agent-Controlled Forgetting, a novel approach to managing context for tool-using agents. This method involves replacing tool results with short summaries while archiving the originals for potential recovery, significantly reducing resource consumption. In an OpenTelemetry debugging case, the method reduced API costs by approximately 70% and decreased input token usage by 50%. While effective in certain scenarios, the approach showed increased requests and decreased efficiency in
News
· 2026-10-09
arXiv has introduced a novel privacy-preserving framework for anomaly detection in vehicle Controller Area Networks (CAN). The framework utilizes a lightweight Transformer encoder to capture temporal signal evolution and a federated learning mechanism that allows multiple vehicles or ECUs to collaboratively improve a shared model without sharing raw CAN data. Experiments demonstrate that this approach maintains efficiency and privacy while delivering robust anomaly detection performance, offerin
· 2026-10-09
Microsoft's research team has released ThinkingBox, a new tool for evaluating AI agent performance in repeated tasks and their impact on backend database states. ThinkingBox simulates 507 business workflows across five domains, testing nine AI models over 20 independent attempts, and introduces three key metrics: pass@1, pass@20, and all-20. The study reveals a significant disparity between single-attempt success rates and repeated success rates, with many seemingly successful failures failing t