Naive Prompt Optimization Outperforms Complex Search Using Teacher Reasoning

NPO utilizes a lightweight single-lineage method to iteratively revise prompts using teacher model rollout feedback. It achieves performance comparable to or better than GEPA with fewer rollouts, demonstrating that stronger teacher reasoning can substitute for complex optimizer-side search.