Tag
Qwen发布QwenGyre，一个面向超长时序（XLong）智能体训练的弹性强化学习框架，包含弹性调度器与轨迹处理器，在Qwen3.8-2.4T-A95B上将NL2RepoBench从52.5%提升至58.5%，并缩短约44%的端到端训练时间。
SpireBench introduces a long-horizon benchmark in which agentic LLMs play the full commercial deck-builder Slay the Spire. Frontier models beat heuristic baselines and match average human players but fall far short of experts, struggling with per-fight play, risk assessment, and cross-run memory.
EvoWeb introduces a benchmark of 320 cumulative requests across 32 web applications to evaluate coding agents on iterative, long-horizon development of evolving web apps, revealing cross-request integration as a key challenge for frontier models.
提出 Recursive Environment Scaling (RES) 框架，在保持任务目标不变的前提下递归变换基准环境以提升难度，既改善长程智能体评估，也使 SFT 在九组模型-基准对比中全部优于原始轨迹微调。
提出了 Keyframe Mnemonics，一种自监督关键帧发现方法，用于解决非马尔可夫环境下的行为 cloning 长时程依赖问题，在合成记忆域达到 100% 成功率，并在机器人操作基准上比最强基线平均绝对提升 13.9%。
提出 Plan-and-Patch 框架，利用扩散语言模型生成结构化计划并通过局部修补（保留前后缀、重新生成受影响区域）实现高效计划修复；在 Natural Plan 上无任务训练时修复成功率约为自回归模型的两倍，并将计划生成延迟降低 39-46%。
提出了 TerminalHorizon，一个用于长程终端智能体的全自动数据合成引擎，构建 1.5K 任务环境并对 Qwen3.5-27B 进行监督微调，在 12 项基准上取得显著提升。
Harvey 探索将经典的 wake-sleep 学习范式应用于长时程法律智能体：在线阶段（wake）执行任务并生成轨迹，离线阶段（sleep）蒸馏可复用经验教训。在 110 项法律任务上，held-out 全通过率从 2.9% 提升到 15.7%，且检索式记忆可将单任务成本减半。
论文提出 Stateless Language Agents (SLAs)，采用「有状态搜索 + 无状态代理」原则，由框架持有研究状态并为每次调用重建专属上下文，在长达十亿 token 的预算下于软件工程、内核优化和算法设计任务中均取得最佳结果，且比最强基线节省超过 84% 的 token。
The paper introduces MIRA, a hierarchical meta-reasoning architecture for long-horizon research agents that separates research allocation from execution, and MIRA-AC, a generative actor-critic that learns credit assignment at meta-reasoning decision boundaries, improving theorem proving and neural architecture research.
The paper introduces AgentMonBench, a benchmark (SpecGAP, SilentSwap, FeedbackTrace) evaluating LLMs' ability to identify consequential agent decisions and requirements-behavior misalignment for evidence-grounded oversight of long-horizon AI agents.
HLA-WM提出一种训练免费的混合线性注意力框架，通过几何引导的检索与递归线性状态重组合来缓解GatedDeltaNet在长时视频世界模型中的远距离遗忘问题，在60秒基准上以12倍内存节省同时提升所有一致性与相机控制指标。
AutoCompact trains coding agents to decide when to compact context, what working state to preserve, and how to resume, using judge-corrected trajectories for SFT followed by RL with task-success rewards, yielding +9.2% on SWE-bench Verified and +5.0% on SWE-PolyBench Verified.
ReLiveGym introduces a diagnostic evaluation environment for long-lived LLM agents that act sparsely over simulated weeks of replayed real-world news, market, and social-media streams, showing that action-timing mechanisms and feedback use are key design axes across eight base models.
论文提出 Legal Research Bench（LRB），一个由专家撰写的 413 题开放式美国法律研究基准，并评测 13 个前沿模型的智能体；最强的 Claude Opus 4.8 在全通过率上仅达 42.9%，表明法律研究智能体仍远未达到可靠水平。
Meta Superintelligence Labs introduces agentic meta-reasoning, an inference-time harness where a controller decides what work to run next in long agent runs, boosting ProgramBench from 63.7% to 71.5% with GPT-5.5 under the same workers and budget. Gains of 3.6-4.2 points hold across ProofBench, ARC-AGI-2 and LongCoT-mini when averaged over frontier models.
The paper proposes Recursive Self-Rewrite (RSR), a framework where a single base model (Qwen-3.8-27B) discovers successful solutions under diverse harnesses and reconstructs them into training trajectories under a general harness via a planner-critic-executor loop, yielding significant pass@3 gains on terminal-bench tasks after supervised finetuning.
EditHero is the first benchmark for long-horizon, part-level 3D editing with natural-language instructions and target images, comparing non-agentic regeneration methods against LLM/VLM agents that edit meshes via code. Non-agentic methods often miss requested changes and disturb fixed regions, while LLM agents follow instructions better and preserve unedited parts, at the cost of slower edits.
The paper introduces PoS, an inference-time framework that builds and continually maintains explicit belief states as LLM agent decision context, validating consistency and detecting 'Belief Trapping' to trigger tailored recovery, achieving top performance across four benchmarks with three LLM backbones.
Dyna Robotics has released Dyna-2.1, a semi-humanoid robot designed for reliable long-horizon autonomy, optimizing for extended operation before human intervention.