Self-refinement improves an LLM output through a feedback loop: the model generates an initial response, critiques it, and revises it. Repeat the cycle until the response meets a desired quality standard. The process requires no additional training data or reinforcement learning. section 5.5.2
Use iterative refinement when:
- A task has complex or hard-to-define goals, such as rewriting a review or optimizing code. section 5.5.2
- An AI response needs to be reshaped into structured data through multiple query rounds; this can improve results but consumes more resources and tokens. section 4.3.2
- A long document is summarized chunk by chunk, refining the current summary as each new chunk is processed. Poor early summaries can propagate errors. section 7.1.2
