GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning

Source: arXiv Authors: 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, Alexandros G. Dimakis, Ion Stoica, Dan Klein, Matei Zaharia, Omar Khattab Date: 2025-07-25 Primary category: cs.CL All categories: cs.CL, cs.AI, cs.LG, cs.SE

Abstract / key passage

GEPA argues that prompt and program optimization should learn from full language-rich trajectories rather than mostly from sparse scalar rewards. It samples reasoning and tool-use traces, reflects on failures in natural language, proposes prompt updates, and preserves complementary candidates on a Pareto frontier instead of collapsing search to one global winner. The paper reports stronger results than several RL-style baselines while using many fewer rollouts.

Harness takeaway

GEPA is interesting for long-lived harnesses because those systems already emit the traces it wants to learn from: reasoning steps, tool calls, outputs, and evaluator feedback. It is therefore best read not as “RL is dead” theater, but as a strong recent case that reflective evolution over durable prompt/program artifacts can be a better fit than scalar-reward RL when rollouts are expensive and traces are rich.