superred-optimizer-gepa

A GEPA (Genetic-Pareto) reflective prompt evolution optimizer for the superred red-teaming framework.

GEPA treats the attack prompt as a single evolving candidate. Each superred run rolls out one candidate; at the end of a run, a separate reflection LM call inspects that rollout (score, target reply, evaluator feedback) and proposes a revised candidate for the next run to try. Across many runs this grows a small pool of candidates by reflective mutation rather than gradient-free search or manual prompt engineering.

Reference: Agrawal et al., "GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning," arXiv:2507.19457 (ICLR 2026). The reflective meta-prompt is reproduced verbatim from the official gepa-ai/gepa reference implementation (MIT-licensed) — see ASSUMPTIONS.md for the exact source files and every deliberate departure this superred port makes (all forced by superred's single-Goal, single-instance session model, where GEPA's paper assumes a full training set to batch and Pareto-select over).

For an agent-target variant of this same reflective loop (surface discovery, tool-return injection), see superred-optimizer-gepa-agentic.

Install

pip install superred-optimizer-gepa

Usage

from gepa_optimizer import GEPAOptimizer

optimizer = GEPAOptimizer(max_attempts=20)

Wire it into a Controller like any other superred optimizer — see the framework's architecture overview and a target's README (e.g. superred-target-chatbot) for a full Controller.run() example.

Behaviour notes

• One run = one rollout of one candidate. The seed candidate is the adversarial goal's own description; each later candidate is proposed by the reflection LM from the best-scoring candidate's most recent rollout.
• Runs in all four adversarial information-access settings without a setting flag. Settings without a score signal collapse candidate selection to latest-wins on ties; settings with feedback drive score-based parent selection and success early-stop. max_no_signal_runs (disabled by default) bounds cost in the score-blind settings.
• No acceptance test / single-instance Pareto. The paper's acceptance test replays a candidate on a fixed minibatch before keeping it; superred's single-Goal, single-attempt-per-run model makes that replay impossible, so every rolled-out candidate enters the pool and parent selection collapses to current-best. Documented as a deliberate departure, not an oversight.
• Reflection parse failures (malformed or missing fenced-code-block output) do not crash the run — the optimizer logs a warning and skips that mutation, keeping the previous best candidate as the parent.

License

MIT for this port's code. The reflective meta-prompt text is reproduced verbatim from gepa-ai/gepa (MIT License, Copyright (c) 2025 Lakshya A Agrawal and the GEPA contributors) — see the attribution comment in src/gepa_optimizer/prompts.py and ASSUMPTIONS.md.