CAS and Meituan Release Survey on Recursive Self-Improvement in LLM Agents

A multi-institution team led by the Chinese Academy of Sciences' Institute of Automation and Meituan, with Shandong University, Renmin University of China and Shanghai Jiao Tong University, has released a survey on recursive self-improvement (RSI) for LLM agents. It frames the field around three questions (when to change, where to change, how to change) and a seven-step lifecycle, organizing 40 representative studies along four lines: prompt optimization, memory evolution, skill evolution and harness evolution, with closed-loop comparisons across 20 of them. The central claim is that agents can keep improving without weight updates by turning interaction experience into validated structure.