Having the model check its own work before finishing. It catches a useful proportion of mistakes, particularly formatting and instruction-following errors, and it is unreliable on the errors that matter most, because a model confident enough to state something wrong is usually confident enough to approve it.
Technical view
Prompts the model to critique and revise its output, optionally against a checklist or rubric. Evidence is mixed: reliable gains where an external signal is available, such as test results or a verifier, and much weaker gains on pure self-assessment, where models tend to ratify their own answers. Effectiveness improves with a separate critic instance, a different model, or a fresh context that does not contain the original reasoning.
Related
Part of AI Terminology Index, section: Agents and Tools.