GEPA optimizes prompts for arithmetic word problems

GEPA improves prompt performance for arithmetic word problems by evolving multi-component prompts with structured feedback. A weak seed prompt is refined using GPT-4o-mini for task solving and GPT-4.1 for reflection. Feedback from failed attempts is fed into GEPA to guide prompt evolution. The optimized prompt outperforms the baseline on a held-out validation set of 100 problems. The process uses a deterministic benchmark with problem types including discount, travel, wallet, and chain scenarios.