OpenEvolve

OpenEvolve is an open-source framework for LLM-driven algorithmic discovery that couples large language models with evolutionary search. Released by codelion / Algorithmicsuperintelligence, it implements the AlphaEvolve paradigm of using LLMs as mutation operators over candidate solutions, with an island-based population model and a CVT-MAP-Elites archive that maintains diversity while tracking per-cell quality. The framework is widely used as a baseline for self-improving-agent research, including in Meta^n and in follow-up work comparing evolutionary to recursive-self-improvement approaches.

Unlike AlphaEvolve (Google DeepMind, May 2025), OpenEvolve is fully open-source under Apache 2.0, written in Python, and supports plug-in replacement of the LLM backbone, the evaluator, and the search algorithm. It has been extended in CodeEvolve (arXiv 2510.14150, October 2025), which adds inspiration-based crossover, meta-prompting, and depth-based refinement on top of OpenEvolve's MAP-Elites archive, and is used as the evolutionary baseline in Meta^n (arXiv 2608.24735).

OpenEvolve is adjacent to evolutionary-algorithm (the classical population-based optimization it implements) and to dgm-darwin-godel-machine (which uses a similar evolutionary archive to search over self-modifying agents). It differs from Gödel Agent in being purely code-evolution-based — no LLM-as-self-modifier — and from Meta^n in not implementing multi-layer recursive depth. The framework is one of the canonical baselines for any new self-improving agent system as of 2026.