Tag
A new frontier-level, long-horizon benchmark called GDPx, the successor to GDPval, has been released; even the best models only score around 12%, highlighting a large gap in agentic task performance.
The paper introduces BABA-Bench, a benchmark of 47 interactive patient-advocacy worlds that checks each agent action in real time for permissions, evidence, and prohibited steps. Evaluating 14 frontier models across 1,974 runs, it finds agents often reach correct outcomes via invalid routes—e.g., Claude Opus 5 hits all outcomes in 81.6% of runs but passes all procedural checks only 52.5%—and that supplying a verified route boosts pass rates to over 91%.
CodeSpecWeave is a framework for synthesizing long-horizon coding tasks with rich, interdependent specifications, yielding 4,670 training tasks and a 100-task benchmark; RL training on these tasks significantly boosts DeepSWE-v1.1 performance over baselines.
Google 发布 Gemini 4 Carbon，被称为其最强的前沿模型，具备最先进的网络安全与长时程任务能力。
Google 发布新一代旗舰模型 Gemini 4 Carbon，具备顶级网络安全与长时程任务能力，在 Artificial Analysis 智能指数 v5.0 上排名第一，并将在网络安全防御领域率先向可信合作伙伴灰度开放。
Browser-Use partnered with Carnegie Mellon researchers to release Odysseys2, a hard long-horizon live web-agent benchmark built from real user tasks, where the best SOTA agent completes only 47/100 tasks.
A UC Berkeley study finds agents that periodically drop their history and restart from their best result outperform parallel agents by up to 15%; the piece also cites Microsoft's Fork-and-Flush and Drew Breunig's 'context inertia' to frame the open question of what state should persist across a context reset.
HaMo is a Harness-Model collaborative framework that combines rapid retrieval of strategy-level experience cards with gradual parametric adaptation to improve long-horizon web agents' retention and cross-task execution, outperforming harness-only and model-only approaches under domain shifts.
HiP3R 提出一种层级式位姿图方法，通过先优化块级 Sim(3) 图建立全局布局，再固定该布局估计每帧共享的 SE(3) 相机位姿，将局部视频重建合并为全局一致的相机轨迹和 3D 几何，无需重新训练。
Shanghai Artificial Intelligence Laboratory has released Sage-A1.5, a 35B MoE agent model with only 3B activated parameters, focused on long-horizon task execution, parallel sub-agent orchestration, and open scientific discovery. On multiple benchmarks, it outperforms much larger models such as Kimi-K3, GLM-5.3, and GPT-5.6-sol with far fewer parameters, while releasing the technical report and model weights as open source.
The paper presents RiseHarness, a framework from Tencent and multiple universities that iteratively improves agent harnesses for long-horizon terminal tasks while keeping model weights fixed, analyzing successful and failed runs to progressively integrate better execution mechanisms.
该论文提出 TURTLE，一种无监督环境设计方法，通过 proposer、solver 和 feedback generator 自动演化出自我调优的长时程工具使用基准，可在 AppWorld、τ-Bench 等评测中提升任务难度、多样性并揭示更广的失败模式。
Navalia is a framework for Cowork Agents combining verifiable task synthesis from corporate financial reports with long-horizon post-training (SFT + RL), achieving consistent gains over its Qwen3.6-35B-A3B base model across five agent benchmarks.
Office Intelligence introduces research environments, harness, and evaluation for long-running AI agent work in office settings, featuring file systems, terminal tools, documents, spreadsheets, chat, and mail within a simulated reconciliation workflow.
DeepSeek V4.1 Flash 1017 matched Opus 5.5 on long-horizon browser use benchmarks while being 19x cheaper per task, establishing a new cost-performance Pareto frontier.
提出 Source-Judged Compaction (SJC)，一种免训练的后处理方法，通过可校准的判断模型 Jev 为压缩摘要中的每条目评分是否能在环境输出中找到依据，并移除无来源条目，从而提升长时程 LLM 智能体的跨运行可靠性 2.1%–8.9%。
TreeOPSD 提出一种无批评者的树状后验自蒸馏框架，用于长时程智能体强化学习，通过将终局奖励递归回传为回合级信用并生成有界教师分布，在多个基准上显著优于 GRPO。
MMSearch-OPD 提出了一种结合可控数据精炼与联合强化学习和在线策略蒸馏的方法，以提升长时程多模态搜索智能体的能力，通过将简单问题扩展为困难且可学习的训练样本取得显著改进。
DeLoop is a self-supervised preference optimization method that reduces repetitive tool-call loops in long-horizon language-model agents by using the agent's own rollouts as training pairs, cutting looping tasks from 57.3% to 6.1% under context compression and improving success rates on larger models.
EdgeWise 是一个边缘-云协作框架，在本地部署小语言模型（SLM）作为主要求解器，并通过强化学习在关键步骤选择性地将任务委派给云端大模型，从而在 SWE-Rebench 上降低高达 62%、在 Terminal-Bench 2.1 上降低 34% 的推理成本，同时保持与纯 LLM 执行相当或更好的性能。