97-page survey by LSTM father maps self-improving agents: memory evolution and tool creation

大模型之路 · wechat · 2026-08-18

A 97-page survey "Self-Improving Agents: A System-Level Survey", co-authored by the father of LSTM, unifies previously conflated concepts—self-reflection, self-correction, self-play, agentic RL, and skill learning—into one coordinate system, compiling 312 related works.

The survey splits self-improvement into three layers: parameter-level continued training at the bottom, architecture-level search and evolution at the top, and two most practical paths in the middle—memory evolution (turning task gains/losses into reusable experience) and tool creation (agents writing their own small tools). Its thesis: as base-model capability plateaus, the real differentiator is whether the system accumulates experience—the dividend of prompt engineering is peaking.

Practical advice: first define what capability should evolve (tools, strategy, or experience); second, add a yardstick—regression tests on fixed tasks, keep improvements and roll back regressions; finally add guardrails—scope limits, logging of every self-modification, auditability, and sandboxed incremental evolution before granting full autonomy. Risks of runaway recursive self-improvement (RSI) and contaminated self-play data are also flagged.

Original post →

More from coding & agent

coding & agent channel →