The reflection pattern is one in which an LLM first produces an answer, then the same (or another) model critiques that answer and asks for an improved revision. The cycle is usually "generate → evaluate → revise," and when used carefully it reduces Hallucination and improves complex reasoning. Self-Critique and Constitutional AI are close relatives of this pattern. Andrew Ng listed reflection as one of four foundational agent patterns in early 2024, and it frequently sits at the heart of an Evaluation Loop in production agents.
MEVZU N°124ISTANBULYEAR I — VOL. III
Glossary · Advanced · 2023
Reflection Pattern
A two-step agent pattern where an LLM critiques and revises its own output.
- EN — English term
- Reflection Pattern
- TR — Turkish term
- Yansıma (Reflection) Deseni