AI Paper Atlas

Research map / Reinforcement learning / RL from feedback

GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning

Lakshya A Agrawal, Shangyin Tan, Dilara Soylu, Noah Ziems, Rishi Khare, Krista Opsahl-Ong, Arnav Singhvi, Herumb Shandilya, Michael J Ryan, Meng Jiang, Christopher Potts, Koushik Sen and 5 more

ICLR 2026 · Reinforcement learning › RL from feedback · approach: RL for LLM reasoning

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Use GEPA reflective prompt evolution instead of GRPO (24,000 rollouts)Qwen3-8B · six tasks aggregatescore: 48.91 -> 54.85Improved
Use GEPA instead of GRPOQwen3-8B · AIME-2025accuracy: 38.00 -> 32.00Hurt
Use GEPA instead of MIPROv2Qwen3-8B · aggregatescore: 47.84 -> 54.85Improved
Use GEPA instead of MIPROv2GPT-4.1 Mini · aggregatescore: 58.67 -> 65.22Improved
Pareto-based candidate selection instead of SelectBestCandidate (ablation)GEPA · Qwen3-8B 4-task aggregatescore: 54.89 -> 61.28Improved
Pareto-based candidate selection instead of BeamSearch (ablation)GEPA · Qwen3-8B 4-task aggregatescore: 53.95 -> 61.28Improved
Add Merge (system-aware crossover) on GPT-4.1 Mini (ablation)GEPA · GPT-4.1 Mini aggregatescore: 65.22 -> 66.36Improved
Add Merge (system-aware crossover) on Qwen3-8B (ablation)GEPA · Qwen3-8B aggregatescore: 54.85 -> 52.40Hurt
Transfer prompts optimized on Qwen3-8B to GPT-4.1 MiniGEPA · GPT-4.1 Mini aggregatescore: 53.03 -> 62.03Improved
Use GEPA instead of MIPROv2 prompt style (instruction-only)GEPA · prompt lengthtokensImproved
Apply GEPA to Sequential10 NPU kernel agentGPT-4o · NPUEvalmean vector utilization: 4.25% -> 30.52%Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy