LSTM之父最新97页综述:Agent如何真正学会「自我进化」?
TL;DR - A 97-page survey (Jilin University, KAUST, Alberta, IDSIA/Schmidhuber's lab) proposes a unified framework for "self-improving agents," cataloging 312 works and defining when an agent genuinely improves versus merely reflecting at runtime.
- Formalizes an agent as base-model parameters θ plus scaffolding Σ (prompt, memory, tools, routing/scheduling/safety logic); self-improvement counts only when execution-derived feedback persistently modifies θ or Σ, so in-context reflection or self-correction alone does not qualify.
- Splits work into two routes: Foundation Model Improvement (self-generated data, model-based scores/critiques, environment experience fed back via SFT/preference optimization/RL — e.g. Self-Instruct, Constitutional AI, WebRL) and Scaffolding Improvement (prompt rewriting, memory curation, tool creation/repair, architecture search — e.g. TextGrad, Mem0, Voyager, Gödel Agent, Darwin Gödel Machine). Risks cited include error re-training, model collapse, catastrophic forgetting, and reward/world-model hacking.
- Treats "skill" as a serializable, reusable update rather than a fifth component, distinguishing object-level skills (external tasks) from meta-level skills (rewriting prompts, memory, tools) that touch recursive self-improvement; surveys six application domains (software engineering, web automation, games, scientific discovery, embodied robotics, general computer control) differentiated by feedback reliability, cost, and reversibility.
- Argues evaluation must target update trajectories, not static scores — per-round gain under fixed budget, transfer, regression, supervision cost, cumulative safety risk — with generator/evaluator separation to avoid judge-gaming; design principle is "fast-loop exploration, slow-loop consolidation," with agents treated as untrusted programs under layered gating since prompt injection can become persistent once written into memory or tools.