Unknown · od Juli 2025 (Paper/GitHub-Erstveröffentlichung); Februar 2026 (optimize_anything API) · 2× · naposledy 29. 6. 2026
GEPA (Genetic-Pareto) is an open-source framework from researchers at UC Berkeley, Stanford, Databricks and others for optimizing text artifacts such as LLM prompts, code, or agent architectures. It combines evolutionary search with LLM-based reflection on execution traces (instead of a scalar RL reward) and selects candidates via a Pareto frontier to preserve diversity. Originally introduced as a prompt optimizer (paper July 2025, ICLR 2026 Oral), it was extended in February 2026 with the "optimize_anything" API to arbitrary text artifacts. It is available as a Python package on PyPI and GitHub, uses LiteLLM to support 100+ LLM providers, and is integrated into frameworks like DSPy, MLflow, Google ADK, and LangChain.
Vlastnosti
| Lizenz | MIT-Lizenz |
| Plattform | Python-Bibliothek (pip install gepa), Python 3.10–3.14 |
| Preis | Kostenlos (Open Source), Kosten fallen nur für genutzte LLM-API-Aufrufe an |
| Protokoll-Kompatibilität | Integrationen mit DSPy, MLflow, LangChain, Google ADK, Pydantic AI, Comet ML Opik |
| Release-Datum | Erstveröffentlichung Juli 2025 (arXiv-Paper); optimize_anything-Erweiterung 18. Februar 2026 |
| Unterstützte Modelle/Provider | Über 100 LLM-Provider via LiteLLM (u.a. OpenAI, Anthropic, Google, Azure, AWS) |