Reflexion(反思)Reflexion
Reflexion
让 agent 在失败后写一份「复盘」,把教训存进记忆,下一轮避免同样的错。Let an agent write a post-mortem after failure, store the lesson in memory and avoid the same mistake next time.
一次任务失败后,模型根据轨迹和反馈生成一段自我反思。
反思被写入长期记忆,下一次尝试时作为额外上下文读回来。
它不改模型参数,只改下一轮的提示与策略。
After a failed attempt, the model writes a reflection over the trajectory and feedback.
The reflection is stored in long-term memory and read back on the next attempt.
No weights change; only the next prompt and strategy change.
agent 最常见的问题是重复犯错。反思把一次失败变成可复用的经验,而不是重新从零试。Agents often repeat the same error. Reflection turns one failure into reusable experience instead of starting from zero.
去 Agent 图谱:看有哪些装得上的客户端、运行时与工具Agent atlas: which clients, runtimes and tools you can install ↗
回到名词地图看第3站的其他 66 个词 →Back to the map: the other 66 terms in station 3 →
- Reflexion 论文(arXiv)Reflexion: Language Agents with Verbal Reinforcement Learning (arXiv 2303.11366) ↗