Survey maps recursive self-improvement in LLM agents
机器之心 reports that researchers from the Institute of Automation at the Chinese Academy of Sciences, Meituan, and other institutions published a survey on recursive self-improvement in LLM agents. The survey organizes the field around when, where, and how agents should update, and maps Prompt, Memory, Skill, and Harness evolution through a seven-step lifecycle. It covers 40 representative studies and compares 20 concrete update loops.
The review provides a shared framework for understanding how agents can turn interaction logs into verified and reusable updates without changing their underlying model weights.
机器之心 2026-11-15 10:29 辽宁
LLM Agent 递归自我改进综述
从 Prompt 优化到 Harness 进化,梳理 RSI 的发展趋势,用三个问题、七步生命周期,理解 Agent 如何把经验转化为持续改进。
想象这样一个场景:一个编程 Agent 调用 API 时遇到报错,发现接口参数已经更新。经过几轮尝试,它终于修复了问题。
#Institute of Automation, Chinese Academy of Sciences#Meituan#Shandong University#Renmin University of China#Chinese Academy of Sciences#Shanghai Jiao Tong University#Li Auto#Huazhong University of Science and Technology#Nanyang Technological University#LLM Agent Recursive Self-Improvement Survey