News
· 2026-10-05
MiniCorp is an AI-driven office simulator designed to address the data scarcity challenge in enterprise-level AGI training. By creating a virtual enterprise environment that models external market dynamics and internal agent decision-making, MiniCorp generates large-scale enterprise data to support agent training and evaluation. The simulator can replicate complex market interactions and allows agents to iteratively optimize their decisions across multiple rounds. Its key strength lies in genera
News
· 2026-10-05
Hugging Face's research team has released a study on the decision-making behavior of tool-using agents, focusing on whether agents can stop relying on ineffective tools based on their judgments of the results. The study found that while agents can accurately judge the results as useless, they rarely stop using the tool based on this evidence. By separating the agents' judgments from their actions, the research reveals a weak dependency on evidence in decision-making and proposes methods, such as
News
· 2026-10-05
Hugging Face's research team introduces PrisMem, a novel approach for agent memory self-evolution. By adopting a capability-driven evolution strategy, PrisMem extends search guidance from overall performance to individual capability dimensions, preserving promising revisions and expanding exploration beyond holistic evolution. Experimental results demonstrate PrisMem's effectiveness, outperforming the strongest baselines by 10.54 and 7.83 percentage points on BEAM-1M and LongMemEval-M benchmarks
News
· 2026-10-05
Hugging Face introduces ACG-Bench, a novel benchmark designed to evaluate the compositional generalization of skills in dual-arm robotic collaboration. The benchmark encompasses 23 task-condition pairs across 8 task families, featuring 6 in-domain conditions and 17 unseen compositions that test reordering, synchronization, combination, and cross-task composition. Utilizing π_{0.5} as the vision-language-action (VLA) backbone, ACG-Bench compares various data augmentation and architectural strateg
News
· 2026-10-05
Hugging Face has introduced RACE (Reliable Action-Chunk Extension), a novel framework designed to address the unreliability of long action sequences in robotic manipulation. By predicting the timing of subskill transitions and conditioning action generation on these predictions, RACE reduces errors at subskill transitions, enabling robots to execute longer action sequences reliably and reduce idle time caused by stop-and-go execution. In simulation benchmarks, RACE outperforms fine-tuning at the
News
· 2026-10-05
Hugging Face has introduced InterMimicGen, a self-evolving motion-imitation framework designed to enhance humanoid robot locomotion and manipulation. The framework consolidates human motion-captured data into humanoid robot references, trains a physics-based generalist tracker, and employs a data flywheel mechanism to iteratively refine motion quality and task completion. Experiments demonstrate its effectiveness in cross-robot configuration retargeting, broad tracking with a single policy, and
News
· 2026-10-05
TikTok has launched a conversational AI-powered Shopping Assistant designed to help users discover and purchase products. The assistant guides users through product browsing via intelligent conversations and offers a one-click checkout feature, streamlining the e-commerce process. This move signifies TikTok's deeper foray into the e-commerce space, leveraging AI to enhance user shopping experiences and boost platform monetization.
News
· 2026-10-05
Apple Machine Learning Research has released a study on user-authored personal sensing systems, focusing on how design tools can empower users to influence and negotiate the ontological boundaries of AI systems. The research utilizes the 'Wizard of Oz' technique to simulate the experience of training personalized machine learning models, aiming to address the limitations of current systems in defining ontological boundaries and provide new insights into AI system design.
News
· 2026-10-05
AWS has launched Amazon Bedrock AgentCore Evaluations, a comprehensive framework for assessing the performance of multi-agent systems. This framework offers three layers of evaluation: built-in evaluators for general quality, custom evaluators for business-specific accuracy, and explainability evaluators for decision transparency. The solution supports both on-demand benchmarking and continuous production monitoring, enabling enterprises to ensure the reliability and explainability of their mult
News
· 2026-10-05
AWS has launched Agentic Retrieval, a feature built on LangChain and Amazon Bedrock Managed Knowledge Base, to address the limitations of traditional single-shot retrieval in handling complex multi-part questions. By decomposing queries into sub-queries, evaluating the sufficiency of evidence, and re-querying when necessary, Agentic Retrieval significantly enhances the quality of answers for multi-hop questions. This article provides a comprehensive overview of the implementation process, API us
News
· 2026-10-05
Safeworld is developing AI-driven digital human technology to ensure robots operate safely in the real world by simulating human behavior and decision-making processes. This approach aims to provide robots with safe behavioral guidance, reducing potential risks to humans. This innovation not only opens new avenues in robot safety but also offers new possibilities for the integration of AI and robotics.
News
· 2026-10-05
LlamaIndex has launched LlamaParse, an Agentic OCR framework designed to address the limitations of traditional OCR in handling complex documents. Agentic OCR employs adaptive reading strategies, self-verification mechanisms, and clear execution objectives to intelligently parse and correct documents. This framework enhances the accuracy of document parsing and provides AI agents with a more reliable data foundation, pushing the boundaries of document processing technology.
News
· 2026-10-05
Independent researchers have identified an AI agent swarm operating on Tencent's infrastructure and targeting Alibaba's map service, Amap. This discovery has sparked discussions about the potential implications of AI agents in commercial competition, cybersecurity, and data privacy. The agent fleet exhibits complex task execution capabilities and dynamic strategy adjustments, distinguishing its behavior from known AI applications.
News
· 2026-10-05
Safeworld is developing digital humans powered by AI to ensure that robots do not cause harm to real humans. This technology provides safe behavioral guidance to robots by simulating human actions and decision-making processes, thereby reducing potential risks. This innovative approach not only opens new avenues in robot safety but also offers new possibilities for the deep integration of AI and robotics.
· 2026-10-05
Luke Salamone's team has distilled the Stockfish chess engine's value function into a ResNet/ViT neural network model, trained on 1 billion chess positions from the Gigafish dataset. The project also releases a full dataset of 3.9 billion positions on Hugging Face, providing a valuable resource for AI chess research. The approach leverages the strengths of CNNs and Vision Transformers to achieve efficient value function approximation in depth-limited search scenarios, demonstrating new potential
News
· 2026-10-05
arXiv has introduced THPL, a novel generative decision support framework tailored for precision feeding in Recirculating Aquaculture Systems (RAS) for rainbow trout. The framework leverages trajectory extraction, hierarchical behavior encoding, and fine-tuning of a Large Language Model (LLM) via LoRA and counterfactual multimodal Direct Preference Optimization (mDPO) to enhance decision accuracy. Experimental results demonstrate that THPL significantly outperforms traditional methods, achieving
News
· 2026-10-05
arXiv has introduced APDMem, a novel hierarchical long-term memory architecture designed to enhance the efficiency of sparse evidence retrieval in complex conversations for personalized LLM assistants. APDMem employs a progressive disclosure mechanism that adaptively adjusts the retrieval granularity based on query complexity, achieving a balance between cost and fidelity. Experimental results on LongMemEval demonstrate that APDMem achieves strong performance in long-context memory reasoning whi
News
· 2026-10-05
arXiv introduces DeReAct, a novel AI agent architecture designed to address the limitations of traditional ReAct-based agents in action authorization and task completion control. By incorporating two distinct gating policies—a Critic for validating actions and a Context Manager for certifying task completion—DeReAct significantly improves the performance of weaker models. In benchmarks like GAIA and SWE-bench Verified, DeReAct boosts Pass@1 scores by 6.5-7.0 and 4.2-5.2 percentage points for Qwe
News
· 2026-10-05
A new study published on arXiv introduces the Comparative Inference for Tool-use Agents (CITA) framework, which aims to improve decision-making for long-horizon tool-use agents. CITA trains a Comparative Inference Model (CIM) that leverages a Bayesian tool-graph simulator and LLM-based semantic judgments to estimate the long-term value of potential tool invocations. Experiments across multiple benchmarks demonstrate that CITA consistently enhances Tool F1 and task success rates, offering a more
News
· 2026-10-04
Hugging Face has released SERA (Self-Evaluating Recursive Agents), a novel framework designed to address the challenges of task decomposition and subtask evaluation in recursive agents. SERA introduces learnable rubrics that allow agents to autonomously generate task breakdowns and dynamically assess subtask completion, reducing reliance on external supervision. Experiments demonstrate that SERA significantly enhances the performance of recursive agents across multiple benchmarks, showcasing str
News
· 2026-10-04
Hugging Face has released Vela, a vision-language-action (VLA) foundation model for robot behavior modeling. Vela introduces a spline-based compact action representation, motion-dependent temporal support, and a shared action interface for heterogeneous embodiments, allowing for adaptive temporal resolution across motions within a fixed output budget. Unlike traditional fixed-rate action chunk representations, Vela more efficiently captures the continuity and smoothness of robot behavior and dem
News
· 2026-10-04
Hugging Face has introduced CornerPercentile, a novel approach for generating extreme scenarios in autonomous vehicle testing. By representing the adversity of a generated scenario as a percentile in the conditional distribution of future risk, and leveraging a percentile-conditioned joint diffusion model with risk-guided sampling, CornerPercentile enables precise control over the extremity of generated scenarios. Experimental results demonstrate its effectiveness, achieving a 98.75% success rat
News
· 2026-10-04
Hugging Face has introduced CoDance, a novel framework for learning reactive and compliant human-humanoid interactions from video. By converting videos of two human dancers into robot reference trajectories and a moving partner, CoDance introduces multi-link compliance augmentation to transform kinematic demonstrations into force-aware training data. Simulation experiments demonstrate that CoDance enables robots to adapt to partner movements while maintaining a stable locomotion style, and physi
News
· 2026-10-04
Hugging Face has released SheetSage2, a novel framework for music transcription that combines synthetic data, task-specific structured decoding, and autoregressive distillation to significantly enhance the coherence and accuracy of generated scores. Leveraging automatically annotated MIDI data for scalable supervision, SheetSage2 integrates complementary musical cues and their temporal dependencies through task-specific decoders, resulting in musically coherent scores. Across eight benchmark col
News
· 2026-10-04
Hugging Face has unveiled Mobile-4DGS, a lightweight unified framework designed for high-fidelity real-time static and dynamic 3D/4D Gaussian Splatting on mobile platforms. The framework introduces a Monte Carlo Specular Energy Aggregator, an Attribute-Conditioned SH Enhancement module, and a Multi-View Alpha-Based Densification and Pruning strategy to reduce storage and rendering overhead while maintaining multi-view consistency. Additionally, its novel 4D explicit representation and Depth-Orde
News
· 2026-10-04
Hugging Face has introduced EVISKILL, an evidence-driven framework designed to enhance continual skill evolution for AI agents. By organizing execution observations into Replayable Evidence Cards and linking edits to their supporting contexts, EVISKILL ensures precise verification and optimization of skill updates. Experiments across three interactive benchmarks and six LLM backbones demonstrate its effectiveness in improving learning efficiency and decision-making for AI agents, marking a signi
News
· 2026-10-04
Hugging Face introduces AutoSciBench, an innovative framework for automatically generating and iteratively adapting benchmarks for evaluating scientific agents. The framework defines tasks through high-level concepts and low-level recipes, using agent trajectories and feedback to refine task generation. This approach addresses the saturation issue of existing benchmarks as agent capabilities evolve. Experiments across computational biology, materials science, and clinical imaging demonstrate tha
News
· 2026-10-04
Hugging Face has introduced DiVeR, a novel approach to enhance decision-critical verifier learning for Vision-Language-Action (VLA) policies. DiVeR estimates decision criticality based on the dispersion of sampled action representations and reweights the verifier learning process towards states where action selection is most consequential. This method significantly improves task success rates without requiring step-level annotations or additional environment interactions. Experiments across LIBE
News
· 2026-10-04
Hugging Face has introduced UndoBench, a novel benchmark designed to evaluate AI agents' task competence and recovery capabilities in the face of operational faults. Spanning 36 base workflows and 36 fault scenarios across 8 enterprise domains, UndoBench decouples task completion from recovery through counterfactual paired trials. The study reveals that while nominal competence reaches 83.54%, the conditional recovery success rate (CRSR) drops to 46.72%, highlighting critical vulnerabilities in
News
· 2026-10-04
Hugging Face has released SearchJev, a novel model designed to enhance the decision efficiency and reliability of search agents. By decoupling System 1 (fast decision-making) from System 2 (complex reasoning and generation), SearchJev enables rapid scoring of legal options based on search states and decision schemas, eliminating the latency and unreliability associated with generative language models. The model leverages Soft-Label Learning for Calibrated Decisions (SLCD) to learn decision proba