Reflection loop
LLM evaluates its own output, identifies errors, generates a new attempt — repeats until satisfied. "Reflexion" paper Noah Shinn 2023.
Self-improvement without retraining. Used in agent frameworks: AutoGPT, BabyAGI, OpenAI o1 (chain-of-thought with reflection during inference). Risk: confidently wrong (LLM thinks correct answer is wrong), or infinite loop. Often combined with external verifier (tests, compiler, calculator).