Self-Refine: The Refinement Loop
Intermediate
The problem, from first principles
Self-Refine (Madaan et al., 2023) has a single model produce a draft, critique it, and revise it, repeating until the critique is satisfied or a budget runs out. No extra training is needed.
From theory to code
Implement refine_loop(generate, feedback, refine, steps), running the draft, feedback and revision cycle.
Constraints
- The feedback value 'OK' ends the loop.
Hints
Hint 1
Generate once, then in each round check the feedback; stop on 'OK' or when the rounds run out, otherwise refine with the feedback.
Keywords
research-papersagentsself-feedbackiterative
Python
Run your code to see results
No results yet
Run your code to check it against the examples, or submit to run every test.