Critique-revise
Two-step process: LLM generates response → second LLM (or same with different prompt) critiques → first LLM revises. Core idea in Self-Refine (Madaan 2023).
Differs from Reflection: split into separate roles. Used in Constitutional AI (Anthropic), debate-based alignment, code review agents. Risk: over-revision degrades quality after 2-3 iterations. Typically combined with stopping criterion (test pass, score threshold).