← AI Terminology
Reflection / Self-Reflection
Reflection (self-reflection) is a pattern where an agent critiques its own draft, plan, or tool outcome and revises before finalising.
It is a simple loop that improves quality without new weights.
It is a simple loop that improves quality without new weights.
Why It Matters in AI
First drafts are often wrong; a second pass that checks errors catches many failures. Reflection prompts, Reflexion-style verbal RL, and reviewer agents are standard in agent frameworks and coding tools.
Key Points
| Aspect | Description |
|---|---|
| Cost | Extra tokens/latency per improvement |
| Loop | Act → observe → critique → revise |
| Risk | Confident wrong critiques; infinite loops |
| Forms | Self-critique prompts, separate reviewer model, unit-test feedback |
| Papers | Reflexion; self-refine literature |
| Related | Verifier models, human-in-the-loop review |
Simple Analogy
Writing an email, reading it aloud, frowning at a confusing sentence, and rewriting — same author, second look.
Common Usage Examples
- Reflexion agents storing verbal feedback
- Code agents: run tests → fix → re-run
- “Critique your answer, then improve” prompts
- Multi-agent author + editor pairs
Summary
In short: Reflection is the act–critique–revise loop where agents improve their own outputs before committing — cheap quality gains at inference.