自我导向修复：让智能体从错误中习得持久教训
原标题：Don't Forget What You Just Learned: Self-Directed Repair Turns Agent Mistakes Into Lasting Lessons
AI 摘要
该论文提出 Self-Direction 机制，让智能体在任务失败后反思自身轨迹并提取教训，将其蒸馏为轻量级 LoRA 修复适配器，存入持久化教训库，并通过路由机制在后续相关任务中自动复用。实验基于 Qwen 系列开源模型，在 GAIA、Claw-Eval 和 Terminal-Bench 2.1 上分别提升成功率 17.0、8.9 和 13.8 个百分点，且能显著减少重复错误、提升纠错可靠性，少量适配器还可泛化到新任务。
正文节选
Don’t Forget What You Just Learned: Self-Directed Repair Turns Agent Mistakes Into Lasting Lessons Abstract Intelligent agents should learn from their mistakes so that experience leads to continuous improvement. Yet today’s agents recover from a mistake only temporarily: in-context reflections are easily forgotten across sessions and compete for limited context, while full-parameter fine-tuning is expensive and can interfere with existing capabilities. We introduce Self-Direction, a mechanism fo