Self-critique is the act of an LLM grading its own answer against explicit criteria — accuracy, completeness, tone, format — and writing out what's missing. It's the core mechanic inside the Reflection Pattern, where the critique then becomes input for the next generation pass. Anthropic's Constitutional AI work systematically teaches a model to critique itself against a constitution and then revise. It's a cheap, effective quality lever in practice; LLM-as-Judge is essentially the same idea executed by a different model.
MEVZU N°124ISTANBULYEAR I — VOL. III
Glossary · Advanced · 2023
Self-Critique
A technique in which a model evaluates its own output against explicit criteria to surface errors and weaknesses.
- EN — English term
- Self-Critique
- TR — Turkish term
- Öz-Eleştiri