Prompt optimization doesn't need longer prompts to work better—by systematically diagnosing errors, diversifying solutions, and stabilizing selections, you can get better results with significantly shorter prompts.
ESPO fixes a major problem with evolutionary prompt optimization: prompts getting bloated (3× longer) without accuracy gains. It uses three phases—diagnosing error patterns, generating diverse candidate prompts, and selecting stable ones—to create shorter, better prompts. On seven NLP benchmarks, ESPO beats the previous best method by 3.76 percentage points while making prompts 47% shorter.