Definition
Self-correction has a model critique and revise its own output, ideally fixing errors without external feedback. The empirical story is mixed: models are decent at spotting their own mistakes when prompted, less reliable at correcting them, and prone to second-guessing correct answers.
Episodes covering this
Worth reading next
Papers we haven't done a deep dive on yet, but would recommend on this topic.
- LEGO-Prover: Neural Theorem Proving with Growing Libraries
- Reflexion: Language Agents with Verbal Reinforcement Learning
- Self-Consistency Improves Chain of Thought Reasoning in Language Models
- Self-Refine: Iterative Refinement with Self-Feedback
- DAgger: A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning
- Chain-of-Verification Reduces Hallucination in Large Language Models