Bezen

Your daily digest

Research papers, made simple to understand

The AI and machine learning papers people are reading most, each one rewritten in plain words: the problem, the idea, what was found and why it matters.

AI Agents23 papers, newest first

    • AI Agents
    • Benchmarks & Evaluation

    OpenDS-World benchmark evaluates AI deep search on real Web tasks with 39.0% top accuracy

    OpenDS-World introduces 200 complex real-world Web search tasks requiring multi-step evidence gathering across diverse modalities and formats; the best AI model achieves 39.0% accuracy, while humans reach 21.1%, highlighting persistent challenges in autonomous Web deep search.

    Paper: Deep Search in Real-World Web Environments · Xiaomin Yu, Hongliang Chen, Zhe Huang and 22 others

    152 upvotes on Hugging Face23 Oct 20267 min readRead the article

    • AI Agents
    • Multimodal & Vision

    SpatialScout benchmark tests VLM agents on active 3D object mapping in indoor scenes

    SpatialScout evaluates vision-language models on autonomous exploration and 3D object localization in 80 indoor scenes with 3,140 targets. The best model achieves 41.71% mean IoU, revealing challenges in precise, complete metric mapping through active perception.

    Paper: SpatialScout: Can VLM Agents Actively Explore and Map 3D Scenes? · Zhaoxi Zhang, Hao Liang, Linxuan Wu and 7 others

    129 upvotes on Hugging Face22 Oct 20266 min readRead the article

    • AI Agents
    • Multimodal & Vision

    FocusGUI improves multi-step omni-modal GUI agents by dynamically prioritizing relevant inputs

    FocusGUI addresses Sequential Distraction in omni-modal GUI agents by adaptively compressing relevant perceptual channels into a compact latent injected into a frozen decoder, improving Type Match by +33.2 points and Exact Match by 3× over frozen backbones across four models.

    Paper: Focused Omni-Modal GUI Agents · Fuyao Zhang, Wenwen He, Jiaming Zhang and 6 others

    225 upvotes on Hugging Face21 Oct 20266 min readRead the article

    • AI Agents
    • Multimodal & Vision

    Matrix-Game 4.0 enables real-time playable game worlds with video model rendering

    Matrix-Game 4.0 separates game logic from visual generation using a lightweight engine, a coding agent, and an autoregressive video model, achieving 480p 60 FPS gameplay with 112 ms latency on a single RTX 5090 GPU.

    Paper: Matrix-Game 4.0: Harnessing Video Models for Real-Time Playable Game Worlds · Jihai Zhang, Runjia Qian, Kaiqin Kong and 15 others

    182 upvotes on Hugging Face21 Oct 20266 min readRead the article

    • AI Agents
    • Benchmarks & Evaluation

    LLM agents approach but do not yet match expert-level operations research solutions

    OR-License benchmark tests 11 large language models on 31 real-world OR tasks with client interaction; best model achieves 88.39% feasibility and 0.87 deployment-adjusted quality versus expert reference of 1, revealing gaps in sustained information acquisition and problem formulation.

    Paper: OR-License: Evaluating LLM Agents as Real-World Operations Research Practitioners · Peisong Wang, Feixue Liu, Yize Meng and 16 others

    196 upvotes on Hugging Face20 Oct 20266 min readRead the article

    • AI Agents
    • Robotics & Physical AI

    HexaShell uses coding agents with RGB triangulation to improve robot manipulation on RoboDojo

    HexaShell lets a coding agent control a dual-arm robot via a thin command interface, evolving small tools from failures using RGB camera triangulation. It achieves 39.7% success on 42 RoboDojo tasks, surpassing learned policies, but struggles with tight insertion tasks requiring precise contact.

    Paper: HexaShell: A Thin Harness for Exploring the Limits of VLM Agents as Robot Coding Agents · Hongcheng Gao, Alex Zhou, Jay Zhu and 13 others

    285 upvotes on Hugging Face19 Oct 20266 min readRead the article

    • Robotics & Physical AI
    • AI Agents

    GPT-6-Astra shows broad robotic skills but faces real-world deployment challenges

    GPT-6-Astra achieves state-of-the-art results on 12 robotic benchmarks using in-context learning and extended interaction, but real-world tests reveal costly demonstrations, physical constraints, and the need for new benchmarks reflecting practical embodied intelligence.

    Paper: Has GPT-6-Astra Achieved Embodied Intelligence? · SenseNova, Joint Research Team, Zhongang Cai and 59 others

    240 upvotes on Hugging Face19 Oct 20267 min readRead the article

    • AI Agents
    • Benchmarks & Evaluation

    BizBench evaluates long-horizon agents on valid, constraint-compliant business task execution

    BizBench is a benchmark of 50 real-world business tasks with 789 documents and 601 tools, testing agents on both final success and intermediate compliance; top models like Claude Opus 5 reach only 64.0% strict success, revealing gaps in execution validity and rule adherence.

    Paper: BizBench: Unveiling Hidden Challenges in Deploying Long-Horizon Agents for Real-World Business · Enhan Zhao, Yongqi Tong, Xuejun Yu and 6 others

    169 upvotes on Hugging Face19 Oct 20267 min readRead the article

    • AI Agents

    OpenGCUA enables models to flexibly use GUI, CLI, and code execution for computer tasks

    OpenGCUA defines and implements a generalist computer-use agent that decides what to observe, how to act, and how to organize work across GUI, accessibility, CLI, and API interfaces, improving task scores by 19.6 points and halving steps on OSWorld-v2.

    Paper: OpenGCUA: Toward a Fully Open Stack for Generalist Computer Use · Hanzhang Zhou, Long Li, Quyu Kong and 12 others

    278 upvotes on Hugging Face16 Oct 20266 min readRead the article

    • AI Agents
    • Image & Video Generation

    ATAS framework improves spatial and temporal consistency in agentic video generation

    ATAS compiles screenplays into shot requests with explicit spatial and temporal references, maintaining persistent scene and event memories to improve cross-shot consistency and enable spatial revisits and event replays, outperforming prior methods on multiple benchmarks and human evaluations.

    Paper: Any Time, Any Space: A World-State-Centric Framework for Agentic Video Generation · Weijie Wang, Haomin Zhang, Ye Li and 7 others

    328 upvotes on Hugging Face15 Oct 20266 min readRead the article

    • AI Agents

    Player-1 improves game control by learning from synthetic environments beyond demonstrations

    Player-1 extends computer-use agents with a richer action space and learns from 6,044 curated gameplay demonstrations, then improves via reinforcement learning in 16 synthetic game environments, boosting MineDojo-300 success by up to 27.8 points and generalizing to unseen games.

    Paper: Player-1: Learning to Play from Synthetic Experience · Mingyu Ouyang, Yiran Wang, Siyuan Hu and 9 others

    279 upvotes on Hugging Face15 Oct 20265 min readRead the article

    • AI Agents

    Jeva architecture enables stable, scalable, and cacheable system one decisions

    Jeva introduces a decision-native model that independently encodes requests and options, refines request state conditioned on options, and explicitly models null outcomes. It achieves strong accuracy, perfect permutation invariance, scalable candidate handling, and large latency reductions via caching.

    Paper: Jeva: System One Models from First Principles · Qi Li, Xinchao Wang

    267 upvotes on Hugging Face14 Oct 20266 min readRead the article

    • AI Agents
    • Multimodal & Vision

    VideoJev unifies video event localization and QA with efficient direct scoring

    VideoJev adapts pretrained multimodal LLMs to directly score temporal intervals and answer options for video understanding, improving accuracy by up to 10.8 mIoU points and speeding inference up to 100× on multiple benchmarks.

    Paper: VideoJev: System One Model for Video Understanding and Interaction · Yuhan Zhu, Xiangyu Zeng, Qingyi Si and 15 others

    196 upvotes on Hugging Face14 Oct 20267 min readRead the article

    • AI Agents

    Agent-Reflex trains language agents to reflect and recover from noisy tool feedback

    Agent-Reflex trains language agents to proactively reflect and recover from failures by mining reflection skills from noisy interactions, fine-tuning on supervised reflection data, and applying on-policy reinforcement learning with reflection-guided distillation, improving robustness on tool-use benchmarks.

    Paper: Agent-Reflex: Teaching Language Agents to Act Reflectively · Guanting Dong, Xinyu Yang, Jiajie Jin and 9 others

    194 upvotes on Hugging Face14 Oct 20265 min readRead the article

    • AI Agents
    • Benchmarks & Evaluation

    AgentSWE benchmark reveals coding agents often fail to build agents meeting user requirements

    AgentSWE evaluates coding agents on creating, editing, and optimizing software agents from natural-language requirements, finding they frequently overfit, produce incomplete runtimes, and overclaim success, with top builders scoring under 31.6 on held-out tests.

    Paper: AgentSWE: Can Coding Agents Build the Agent You Actually Want? · Jiahao Wang, Hongjin Qian, Yuyang Hu and 2 others

    202 upvotes on Hugging Face13 Oct 20266 min readRead the article

    • AI Agents

    Trace2Env simulates environments from interaction traces without original systems

    Trace2Env reconstructs environment knowledge from historical interaction traces into a structured worldbook and uses an agentic language world model to simulate faithful, stateful environment responses, improving next-observation fidelity and long-horizon consistency across nine environments.

    Paper: From Traces to Agentic Worlds: Agentic Language World Models for Interactive Environment Simulation · Quanyu Long, Xiao Chen, Jianda Chen and 4 others

    105 upvotes on Hugging Face11 Oct 20267 min readRead the article

    • AI Agents

    Recursive Self-Rewrite improves model task-solving by rewriting multi-harness trajectories

    Recursive Self-Rewrite (RSR) uses a single model to collect successful task solutions under diverse execution harnesses, then rewrites these trajectories into a general harness format for supervised finetuning, boosting pass@3 rates by up to 20.8 points on complex benchmarks.

    Paper: Scaling Trajectories for Complex Tasks through Recursive Self-Rewrite · Zongxia Li, Yucheng Shi, Zhongzhi Li and 7 others

    103 upvotes on Hugging Face11 Oct 20266 min readRead the article

    • AI Agents
    • Reinforcement Learning & Alignment

    Self-Retrospection Distillation teaches agents to anticipate task demands before acting

    This paper introduces Self-Retrospection Distillation (SRD), a method that uses post-hoc trajectory hindsight to train agents to predict task-relevant knowledge and pitfalls before interaction, improving performance especially when reward signals lack contrast.

    Paper: Self-Retrospection Distillation: Turning Post-hoc Experiences into Prior Foresight · Haoxiang Zhang, Qinglin Chen, Hiroaki Hayashi and 9 others

    146 upvotes on Hugging Face10 Oct 20266 min readRead the article

    • Robotics & Physical AI
    • AI Agents

    MotorMind enables zero-shot robot manipulation using general vision-language models

    MotorMind uses a frozen general-purpose vision-language model to generate mid-level robot actions, with asynchronous monitoring and feedback, achieving 66.7% success on LIBERO-PRO and 95% on a real robot without task-specific training.

    Paper: MotorMind: Scaffolding General Vision Language Models for Zero-Shot Robot Manipulation · Bingxuan Li, Siqi Song, Yizhuo Wu and 3 others

    114 upvotes on Hugging Face10 Oct 20266 min readRead the article

    • AI Agents

    AgentGarten enables evolving agents via code-driven worlds and neural rendering

    AgentGarten couples explicit simulators with a shared neural renderer trained by Adversarial Forcing to create real-time interactive environments. Pretrained agents learn efficiently from rendered observations, improving through playbooks across rounds, achieving tool-use strategies in just a few games.

    Paper: AgentGarten: Code Worlds for Evolving Agents · Jiawei Chi, Shangchen Miao, Zhiyuan Shi and 11 others

    133 upvotes on Hugging Face9 Oct 20267 min readRead the article

    • AI Agents
    • Benchmarks & Evaluation

    Learn2Play Bench tests LLM agents' ability to learn novel rules from interaction

    Learn2Play Bench introduces text-based games with hidden, unfamiliar rules to evaluate how well LLM agents learn from experience without weight updates. Experiments show raw memory retention aids learning, humans outperform agents by exploring more, and agent harnesses impact performance and cost.

    Paper: Learn2Play Bench: How Well Do LLM Agents Learn from Experience in Unfamiliar Environments? · Yibo Li, Jinhang Qiu, Zhi Zheng and 4 others

    108 upvotes on Hugging Face9 Oct 20266 min readRead the article

    • AI Agents

    nanoMuse offers an open-source personal agent running on all user devices

    nanoMuse is an open-source personal agent that runs locally on a user's phone and computer, sharing a conversation across devices via an optional relay, with user-controlled memory and permissions, aiming to replicate and extend Meta's Muse agent in an open, inspectable form.

    Paper: nanoMuse: An Open-Source Personal Agent for Every Device You Own · Guangyi Liu, Yong Liu, Jiangning Zhang

    84 upvotes on Hugging Face8 Oct 20266 min readRead the article

    • Benchmarks & Evaluation
    • AI Agents

    RealCompanion benchmarks AI understanding of real long-term human conversations

    RealCompanion releases ten real AI-human conversations spanning up to 120 days, with detailed annotations linking replies to past messages. It reveals that memory use is rare but often requires distant context, and current AI systems struggle to detect when to recall past information.

    Paper: RealCompanion: Benchmarking Human Understanding from Reasoning over Longitudinal Real-World Conversations · Arman Behnam, Sunglyoung Kim, Liangwei Yang

    262 upvotes on Hugging Face5 Oct 20265 min readRead the article