WTF is Reflection?
An agent reviews its own output or failed attempt, writes down what went wrong, and tries again.
Also called: reflexion, self-critique, self-reflection
TL;DR
Instead of accepting the first draft, the agent critiques it ("the tests failed because I forgot the edge case"), keeps that lesson, and makes a better second attempt. It is a cheap way to improve quality without retraining the model.
How it works
Reflexion formalised this: after a trial, the agent gets feedback (test results, an evaluator score, environment signals), generates a verbal self-reflection, stores it in an episodic memory buffer, and conditions the next attempt on it. Related variants include evaluator–optimizer loops, where a second model or prompt grades the output and the first revises until a bar is met.
Origin
- Coined / introduced by
- Shinn et al.
- When
- 20 Mar 2023
- Story
- Introduced as 'Reflexion: Language Agents with Verbal Reinforcement Learning'.