OpenAlpha_Evolve

About

DeepMind's AlphaEvolve showed that evolutionary search over LLM-generated programs can produce algorithms that beat hand-written baselines, but the system was never released. OpenAlpha_Evolve reconstructs the concept in Python: a PromptDesignerAgent frames the task, a CodeGeneratorAgent writes and mutates candidate programs as diffs, an EvaluatorAgent runs them against tests inside Docker containers, a DatabaseAgent archives the population, and a SelectionControllerAgent drives generations forward. Models route through LiteLLM with Gemini as the default configuration, and candidates can also be inspected through a Gradio web UI. Tasks are defined in YAML examples (shortest path among them), and the whole loop runs from a single python -m main invocation after cloning and pip installing. Researchers and hobbyists in evolutionary computation and LLM-driven program synthesis use it as an accessible AlphaEvolve-style testbed; the repo is experimental and community-maintained.