News
· 2026-10-13
Hugging Face has introduced RegisterWAM, a novel model that enhances robot action conditioning by incorporating learnable non-spatial registers. This innovation suppresses spatial outliers and concentrates high-norm responses, improving the model's ability to learn actions in complex environments. RegisterWAM demonstrates strong performance on benchmarks like LIBERO-Plus (81.64% success rate) and RoboTwin 2.0 (93.0% average success), with notable real-world manipulation improvements.
News
· 2026-10-13
Hugging Face has released GroundGUI-8B, a GUI agent model that leverages the ScreenRef dataset and the Distillation-to-Reinforcement Learning (D2RL) method to significantly enhance multimodal visual-language understanding. The ScreenRef dataset includes 586K GUI screenshots and 633K natural language referring expressions, covering tasks such as text recognition, visual matching, and spatial understanding. The D2RL method addresses the limitations of supervised fine-tuning and the sparse reward p
News
· 2026-10-13
Hugging Face introduces CoSIR (Collaboration via Surprisal-based Information Rewards), a novel post-training approach for enhancing collaboration behaviors in multi-agent systems with large language models (LLMs). CoSIR employs a Bayesian evaluator to score shared information and requests based on reductions in team surprisal, encouraging agents to effectively share evidence, request missing information, and integrate teammates' contributions in complex tasks. Experiments demonstrate that CoSIR
News
· 2026-10-13
Hugging Face has released DeRef, a model designed to address the derivation gap in tool-call argument verification. By generating references from the full history without seeing the candidate, DeRef significantly improves the accuracy of AI agents in evaluating tool-call arguments. Across multiple benchmarks, DeRef boosts AUROC from chance levels to between 98.40% and 99.59%, demonstrating its potential to enhance the reliability of agent decision-making.
· 2026-10-13
Preset.io has released Agor, an AI platform designed for multiplayer AI agent collaboration, addressing the limitation of individual AI agents running in isolation. Agor enables users to operate multiple agents simultaneously, enhancing task efficiency and enabling more complex intelligent interactions. This innovation opens new possibilities for AI agents in team collaboration, task allocation, and real-time interaction scenarios.
News
· 2026-10-13
Hugging Face introduces ACCRUE, a novel continual off-policy reinforcement learning (RL) framework for robot foundation models (RFMs). ACCRUE employs lightweight actor and critic networks to transfer knowledge across tasks, enabling efficient learning of new tasks while retaining previously acquired behaviors with near-zero forgetting. Its key mechanisms include adapting the inherited critic only on current-task transitions and distilling the actor towards cached targets on a small number of ret
News
· 2026-10-13
Hugging Face has released ARBITER, a 35B-parameter Mixture-of-Experts (MoE) agentic verifier designed to revolutionize AI programming by enhancing code selection, verification, and optimization. Trained with reinforcement learning in a generic execution sandbox, ARBITER generates verifiable counterexamples as feedback, significantly improving the accuracy of code verification and the performance of agents in complex programming tasks. In the USACO-JUDGE benchmark, ARBITER increased the mean bina
News
· 2026-10-13
Hugging Face has released a study on the decision failures of System One models, which make rapid, single-pass choices from a set of options. The research, conducted on a benchmark of agent decision states across simulated environments, customer service, tool use, web navigation, and coding tasks, reveals that most incorrect actions stem from correct readings. The more accurate a system is, the more its errors take this form. These errors are not random noise but cluster on specific states, and
News
· 2026-10-13
Hugging Face has released GreenWAM, a world foundation model for robotics that aims to unify the prediction of how robot actions change the scene and the selection of optimal action policies. GreenWAM employs a shared anatomical action space, multi-modal joint training, and a generator-based policy with a split action expert, achieving superior performance in world modeling and control across diverse robotic forms. It demonstrates top-tier trajectory, depth, and 3D accuracy on WorldArena 1.0 and
News
· 2026-10-13
On October 13, 2026, DeepSeek released the Open SWE framework as open-source. Built on Deep Agents and LangGraph, Open SWE provides essential architectural components for internal coding agents, streamlining development workflows and reducing engineering complexity. This framework offers developers an efficient and scalable solution, advancing the application of agent technology in code generation, file management, and system interaction.
News
· 2026-10-13
Hugging Face has introduced SWE-MISBEHAVIOR, the first taxonomy and evaluation framework for diagnosing boundary-violating misbehaviors in coding agents. The framework includes 7.4K human-in-the-loop synthesized samples across three layers of agent action space: Code Generation, File Management, and System Interaction, and defines 11 distinct misbehaviors with a fully automated evaluation protocol. Experiments on state-of-the-art open-source coding agents, such as OpenHands and smolagents, revea
News
· 2026-10-13
On October 13, 2026, DeepSeek launched the beta version of Deep Agents Deploy, an open-source, model-agnostic agent harness designed for fast and production-ready deployment. This framework simplifies the deployment process for AI agents, reducing engineering overhead for developers working on complex AI tasks. It also supports flexible task switching and data persistence, making it a valuable addition to the open-source agent ecosystem and providing a more efficient and cost-effective solution
News
· 2026-10-13
ReTerm is a novel terminal agent designed to improve robustness by learning to recover from its own failed tasks. It leverages a customized data pipeline to create ReTerm-10K, a dataset for Supervised Fine-Tuning (SFT) cold start, and introduces ReTerm-GRPO, an online Reinforcement Learning (RL) algorithm that generates multiple self-recovery attempts from failure states. Experiments demonstrate that ReTerm-9B achieves a Pass@1 score of 34.8% on Terminal-Bench 2.0, outperforming the base model b
News
· 2026-10-13
Hugging Face has introduced Urbex, an agentic vision-language framework for spatial grounding in city-scale 3D scenes. Unlike traditional candidate-ranking approaches, Urbex casts grounding as active spatial search, representing the 3D scene as an interactive multi-view environment. This allows a VLM agent to query landmarks, zoom into regions, render oblique views, and commit grounded bounding boxes. Optimized with reinforcement learning on CityRefer, Urbex achieves state-of-the-art performance
News
· 2026-10-13
RUC-NLPIR has introduced Agent-Reflex, a novel training framework designed to enhance the reflective capabilities of language agents. The framework operates in three stages: Environment Reflection Skill Discovery (RSD), Reflective Supervised Fine-Tuning (ReSFT), and On-Policy Reflective Policy Optimization (ORPO). Evaluated on multiple benchmarks, Agent-Reflex achieved an average score of 58.6 on long-horizon tool-use tasks and 29.1 on AgentNoiseBench, outperforming strong environment-scaling an
News
· 2026-10-13
DeepSeek has released the Deep Agents SDK, featuring advanced context management for long-running AI tasks through offloading, summarization, and filesystem abstraction. This SDK aims to prevent context rot, enhance AI agent performance in complex tasks, and enable flexible task switching and data persistence.
News
· 2026-10-13
Hugging Face has introduced Jeva, an innovative decision-making architecture designed to address the shortcomings of current systems in handling the structural requirements of decision problems. Jeva incorporates four key mechanisms—null-awareness, permutation invariance, option scalability, and cacheability—to deliver precise decision-making in complex scenarios. Its architecture, featuring a shared encoder, independent option representations, and a compact decision stack, significantly reduces
News
· 2026-10-13
Hugging Face has released WAVE-0, a framework that bridges Vision-Language-Action (VLA) policies with World Action Models (WAMs) through a lightweight latent world model. By leveraging the semantic understanding of pretrained VLAs and explicit future prediction, WAVE-0 enhances robotic manipulation planning and execution. The model achieves a 91.47% success rate on RoboTwin 2.0 and 98.7% on LIBERO, demonstrating strong performance and generalization across diverse tasks and distribution shifts.
News
· 2026-10-13
Hugging Face introduces SwitchMem, a dynamic memory modality selection framework designed to enhance the performance of vision-language model (VLM) based GUI agents. Unlike traditional methods that rely on fixed memory representations, SwitchMem dynamically selects between textual and visual memory representations at each interaction step, significantly improving agent performance in complex tasks. Experiments across multiple GUI benchmarks demonstrate that SwitchMem outperforms baseline methods
News
· 2026-10-13
Hugging Face has released EgoExplorer, a 100-hour multimodal observatory for human locomotion and manipulation, built with a wearable system. The system aligns egocentric video with reconstructed whole-body motion, fine-grained finger articulation, and tactile observations, enhanced by the Calibrate-Complete-Constrain pipeline for precise data capture. EgoExplorer demonstrates a 71.9% reduction in body-point error compared to the strongest baseline and shows improvements in hand stability and co
News
· 2026-10-13
Hugging Face has released JevGPT, an innovative dual-system architecture for autonomous research agents. JevGPT enhances research efficiency and coverage by coupling fast evidence processing with deliberative research planning. Its core mechanisms include parallel independent source evaluation and dynamically adjusted research step selection. Experiments show that JevGPT excels in fixed-corpus question answering and document ranking tasks, achieving faster completion and lower costs, representin
News
· 2026-10-13
LangChain has introduced a novel solution that leverages the Deep Agents CLI to automate the migration of inline code into testable, CI-ready examples, addressing the common issue of stale code samples in documentation. This approach transforms static code snippets into dynamic test cases, enhancing the accuracy and reliability of documentation and providing developers with a more efficient tool for maintaining code examples.
News
· 2026-10-13
Google has introduced a new feature in Google Play that allows users to build AI-powered personal Android apps without the need to publish them publicly. This move aims to streamline the development and use of personalized apps while enhancing user privacy and reducing the barriers to entry for private app deployment.
News
· 2026-10-13
Hugging Face has introduced WorkRover, a novel framework for training long-horizon agents through recursive environment evolution. By jointly constructing and validating semantically aligned work contexts, executable tasks, and evidence-grounded verifiers, WorkRover enhances agent performance in complex tasks. The framework incorporates Factorial Compatibility Replay (FCR) and Verifier-to-Turn Credit Assignment (VTCA) to optimize component revisions and provide finer-grained training signals. Ex
News
· 2026-10-13
LangChain has announced a deep integration with ScaleDown and Deep Agents, leveraging ScaleDown's constrained decoding models to enable agents to summarize history, compress tool results, and extract data faster and more cost-effectively. This collaboration aims to address efficiency bottlenecks in long-running agent tasks, providing developers with more powerful tools for building advanced AI systems.
News
· 2026-10-13
Hugging Face has introduced DeltaTTG, a novel reinforcement learning approach designed to address the challenges posed by sparse terminal rewards in multi-turn agent training. DeltaTTG leverages a shared progress meter to learn a reusable progress signal and optimizes the training process using low-cost supervision data. On tasks like ALFWorld, PaperSearchQA, and BFCLv3, DeltaTTG outperforms the existing method GRPO by 5.7 and 7.2 points in average scores, respectively, while maintaining similar
· 2026-10-13
AnythingLLM has announced its migration from Ollama to the Lemonade embedded SDK and llama.cpp. This shift aims to leverage the local deployment capabilities of the Lemonade SDK to enhance the efficiency and data privacy of AI agents, while utilizing the lightweight nature of llama.cpp to optimize resource usage. Post-migration, developers will be able to deploy AI agents more flexibly in local environments, reducing reliance on cloud servers.
News
· 2026-10-13
Industrial Next has developed a technology that allows factory robots to switch tasks in minutes rather than weeks. This breakthrough aims to help manufacturers automate complex and delicate tasks traditionally requiring human labor. By simplifying the robot task configuration process, Industrial Next significantly enhances production efficiency, reduces the risk of human errors, and offers new possibilities for the intelligent transformation of the manufacturing industry.
News
· 2026-10-13
LangChain has enhanced its coding agent's performance on the Terminal Bench, moving from the top 30 to the top 5 by implementing self-verification, tracing, and context optimization techniques. This improvement highlights the potential of engineering optimizations in boosting the efficiency and accuracy of AI agents, offering new insights for agent development.
· 2026-10-13
Tencent has recently open-sourced Agent Memory, an AI agent memory system that runs 100% locally using SQLite. This system enhances data privacy and operational efficiency by eliminating the need for cloud server dependencies for storing and retrieving agent memories. It supports a variety of agent tasks and is optimized for resource-constrained devices, offering developers a flexible and efficient solution for building AI agents.