AREX-2：通过长时程反思任务推进自我改进智能体
原标题：AREX-2: Advancing Self-Improving Agents through Long-Horizon Reflective Tasks
AI 摘要
该论文提出 AREX-2，旨在提升 LLM 智能体的自我改进能力，即测试时通过多轮迭代优化解决方案。研究基于 Qwen3.8-27B 构建智能体，利用机器学习工程和算法编程任务合成长期反思轨迹进行训练。在 MLE-bench Lite 上达到 81.8，Frontier-CS 上 70.7，并迁移至深度研究任务，在 BrowseComp、HLE、GAIA、DeepSearchQA 上分别取得 84.0、52.6、92.2、93.8 的成绩，且随轮次预算增加持续提升。
正文节选
AREX-2: Advancing Self-Improving Agents through Long-Horizon Reflective Tasks Abstract We present AREX-2, an effort to advance the self-improving capability of LLM agents, which we define as the ability to iteratively refine a solution at test time. This ability rests on two complementary capabilities: reflection, which produces a solution better than the current one, and long-horizon execution, which keeps the iteration effective over many rounds. We hypothesize that both capabilities are domai