Self-Correction
Self-Correction enables LLMs to iteratively improve their own responses through a three-stage cycle: Generator creates initial draft, Critic/Feedback evaluates and identifies issues, Refiner incorporates feedback into improved version. The process repeats until quality targets are met. Unlike external feedback, this approach requires no additional training, fine-tuning, or external data — it operates entirely through prompts at inference time. Studies show ~20% quality improvement in human evaluations.
- Radar Signals0