gepa-rlm-exp-20260219-031221

GEPA vs GEPA+RLM prompt optimization experiment on AIME math problems.

Task LM: openai/gpt-4.1-mini | Reflection LM: openai/gpt-5 | Last updated: 2026-02-19 05:07 UTC

Results

Run	Method	k	Val Score	Test Acc	Tokens	Cost	Time
fixed_rlm_k20	rlm	20	40.00%	41.33%	1,805,475	$0.0000	5388s

Learning Curves

Experiment Config

{
  "script_name": "run_experiment.py",
  "model": "openai/gpt-4.1-mini",
  "reflection_lm": "openai/gpt-5",
  "hyperparameters": {
    "task_lm": "openai/gpt-4.1-mini",
    "reflection_lm": "openai/gpt-5",
    "seed_prompt": "You are a helpful assistant. You are given a question and you need to answer it. The answer should b..."
  },
  "input_datasets": [
    "AI-MO/aimo-validation-aime",
    "MathArena/aime_2025"
  ],
  "description": "GEPA vs GEPA+RLM prompt optimization experiment on AIME",
  "num_runs": 1,
  "last_updated": "2026-02-19 05:07 UTC"
}

Dataset Configs

This repo contains multiple configs (subsets). Load them with:

from datasets import load_dataset

results    = load_dataset("reasoning-degeneration-dev/gepa-rlm-exp-20260219-031221", "results", split="train")
traces     = load_dataset("reasoning-degeneration-dev/gepa-rlm-exp-20260219-031221", "traces", split="train")
val_traces = load_dataset("reasoning-degeneration-dev/gepa-rlm-exp-20260219-031221", "val_traces", split="train")
curves     = load_dataset("reasoning-degeneration-dev/gepa-rlm-exp-20260219-031221", "curves", split="train")

Config	Description
results	Per-run summary: method, k, val/test scores, tokens, cost, optimized prompt
traces	Per-example test set traces with model responses and scores
val_traces	Per-iteration validation traces with the prompt used at each step
curves	Per-iteration learning curves: tokens, scores, candidate prompts

State file: state.json

For experiment design details, see EXPERIMENTS-DESIGN.md