GEPA prompt optimization experiment on AIME math problems. Task LM: openai/gpt-4.1-mini | Reflection LM: openai/gpt-5 | Reflection Mode: failures_only | Last updated: 2026-02-19 21:50 UTC

	Results

Run Method k Mode Val Score Test Acc Tokens Cost Time

fixed_rlm_k20 rlm 20 failures_only 46.67% 30.67% 1,150,570 $0.0000 5996s

	Learning Curves





	
	Experiment Config

{ "script_name":… See the full description on the dataset page: https://huggingface.co/datasets/reasoning-degeneration-dev/gepa-rlm-exp-failures_only-20260219-191545.