Tag
Kolab 介绍了一个用于评测人机协作能力的新基准，包含来自40个职业的50位专业人士在115小时以上真实长时程工作流中与AI代理协作的屏幕录像及质量判断，发现前沿模型普遍高估协作质量且难以区分高质量与低质量协作。
LLM-based agents work better with structured data, as demonstrated by Harvey's context graphs for long-horizon legal tasks on the Legal Agent Bench; the author notes DAIR.AI is exploring similar approaches for research tasks.
本文提出 Contextual Credit Policy Optimization (CCPO)，一种无需价值critic的步级信用分配方法，通过上下文相似度加权的同伴比较与未来进度信号，为LLM智能体在长程任务中的强化学习提供更精细的优势估计，在ALFWorld和WebShop上显著优于GRPO。
LHTB 2.0 介绍了一个扩展的长时程终端基准，包含 77 个跨 13 个领域的任务，通过自动化任务构建流水线和密集进度阶梯来衡量 AI 代理在长达三小时工作流中的持续进展，GPT-6-astra 以 0.69 的密集平均奖励领先。
评论介绍了 UCLA 与 Adobe Research 的 Tev-Gym 工作：通过构建可验证的长任务环境与稠密中间进度奖励来训练长时程 Agent，Tev-27B 在 ALE-CM 计算与数学子集上以 Pass@5 比肩 GPT-6 Astra 与 Claude Opus 5，Tev-320B 则将 GLM-5.3-Flash 在 Terminal-Bench 4.0-CM 上的 Pass@1 从 29% 提升至 37%。
SERO 提出通过在真实策略状态上训练可执行的 oracle 续段（并按策略可达性加权）来改进长程 agentic 强化学习，且推理时无需额外开销；在 ALFWorld 与 WebShop 上验证了其有效性。
Doc2Harness 提出了一个两阶段基准，直接评测 LLM 构建长周期编码代理 harness 的能力；在 19 个前沿模型中，GPT-6Astra 以 53.05% 的通过率领先，而表现最弱的模型仅为 27.44%。
The paper introduces TerminalHorizon, an automated data-synthesis engine that generates long-horizon terminal agent trajectories by reconstructing environments around real-world professional tasks, then fine-tunes Qwen3.5-27B to obtain TerminalHorizon-27B, achieving gains across 12 coding, agentic, and reasoning benchmarks.
本文提出层级子任务执行（HSE）框架与 RoboCombine 基准，研究具身操作中技能重组的组合泛化极限，指出失败源于「起始状态差距」，并提出 PRTI 轻量后训练策略以提升跨子任务边界的执行成功率。
ProgAgent 提出一个完整的程序化工具调用（PTC）训练框架，通过可验证的长时程任务合成、轻量级 PTC 运行时和两阶段训练配方，显著优于同规模基线，在 LOCA-bench 64k token 状态下达 30.7% 通过率而基线为零。
A practitioner details the patterns built to keep fleets of LLM agents alive on multi-hour jobs (heartbeats, epoch fencing, intent/done records, parent audits) and explains why failures like credential delivery, workspace resets, and missing dead-man switches still cause silent agent deaths, asking the community for concrete fixes.
TaReD 提出一种工具感知的递归分解方法，通过按功能关系将工具组织成能力层级，让智能体按需发现工具并将复杂长时程任务递归分解为子任务树，实验显示端到端任务成功率最高提升 40 个百分点。
AgentHorizon introduces a benchmark of 1,373 computer-use tasks with human-recorded trajectories to evaluate agentic judges on long-horizon, multi-application tasks. The study finds GPT-5.5 achieves 80.9% balanced accuracy as the best agentic judge, and reveals that tool use helps some models but hurts open-weight models.
pmtui 是一款终端工具，用于管理多个 AI 智能体并处理长时程任务，通过在智能体之外维护循环、每次根据目标和最新状态重建提示词来避免任务漂移。
The paper introduces Trajectory-Anchored Pruning (TAP), the first structural pruning framework for RL-trained long-horizon agents, which couples channel pruning with trajectory-anchored on-policy recovery. With 60% of FFN channels removed, TAP retains 99.2% and 88.0% of dense 7B agents' task success on ALFWorld and WebShop while cutting GPU time per successful task by roughly 22% and 17%.
A practitioner discussion exploring whether long-running AI agent failures stem from model capabilities or from the scaffolding around them, highlighting error compounding, context pollution, and weak self-correction as key failure modes.
The paper introduces ASCENT, a method for online agentic test-time training that lets an LLM agent improve its own weights across tasks by self-distilling verified execution trajectories into persistent LoRA updates, evaluated on ALFWorld, WebShop, and AppWorld.
The paper introduces Generalized Implicit Temporal Abstraction (GITA), a method for offline goal-conditioned reinforcement learning that trains a single k-conditioned value function across multiple temporal abstraction scales, resolving the trade-off between long-range value signal and local resolution. GITA outperforms offline GCRL baselines like HIQL and OTA on OGBench, raising average success rates by 25 percentage points over HIQL.
VeriHarness turns an LLM generator into an agentic verifier—using a workspace, evidence tools, reusable verification skills, a disagreement resolver, and a consensus challenger—to select reliable outputs for long-horizon tasks without reference answers. It achieves top selection scores across five benchmarks, gains of 6.2–6.4 points over single rollouts with Gemini 3.5 Flash and Claude Opus 4.8, and releases ~26,000 rollouts costing over $100,000.
The paper introduces ContextProgress-Bench, a benchmark showing that embodied Progress Reward Models fail at progress estimation in long-horizon tasks without proper context, and proposes ProgressCompass, an agentic loop that supplies the needed context via general-purpose VLMs, cutting progress error by 63-82%.