GEPA

Overview

GEPA is a prompt and program optimizer that learns from full language-rich traces rather than mostly from scalar rewards. It reflects on trajectories in natural language, proposes prompt updates, and keeps complementary candidates on a Pareto frontier.

Why it matters

It matters because real harnesses already emit the traces GEPA wants to learn from: reasoning steps, tool calls, tool outputs, and evaluator feedback. That makes it a much better fit for compound systems than a story built purely around reward signals.

Distinctive trait

Its distinctive trait is reflective evolution with Pareto retention: it preserves different candidates that succeed on different cases instead of immediately collapsing search to one global best prompt.

Relationships

Read GEPA with rlprompt, dspy, reflexion, and prompt-optimization-eval-transfer-robustness-open-questions. It also sits naturally beside self-evolving-workflows, where improvement happens through durable external artifacts rather than weight updates.