Verification · Feedback and evaluation
Reviewer optimization
Tunes reviewer prompts and instructions against the evaluation set instead of by hand.
Sub-techniques
-
Prompt optimization
V1068.001Search prompt variants against a metric.
-
Instruction updates from feedback
V1068.002Update reviewer instructions from accepted and rejected findings.
How to build it
- Define the metric first (from reviewer evaluation).
- Optimize on a train split; report on a held-out split.
- Review every prompt change like code.
Signals in a repository
- Optimizer scripts and versioned prompt files in the repository.