codelion/openevolve
Openevolve
OpenEvolve is an evolutionary algorithm framework that uses large language models to autonomously discover and optimize programming algorithms. It functions as an algorithm discovery engine and code search tool, evolving populations of candidate programs to find efficient implementations and hardware-specific speedups.
The system treats both code and system instructions as evolvable entities, utilizing an automated prompt optimizer to iteratively refine model performance. It maintains search stability through niche-based population management to preserve diversity and employs a closed-loop feedback mechanism that injects runtime errors and logs back into the generation process for autonomous bug fixing.
The framework also includes an agent orchestrator that aggregates responses from multiple model APIs using weighted logic and fallback strategies. To support scientific computing, it implements deterministic execution by managing consistent random seeds across all stochastic components.
The project provides an interactive dashboard for visualizing evolution progress and real-time performance metrics.
Features
- Evolutionary Algorithm Discovery - Implements an evolutionary pipeline using LLMs to autonomously discover and optimize efficient algorithms.
- Evolutionary Code Optimizers - Uses an LLM-driven evolutionary process to iteratively generate and mutate code for algorithm discovery.
- Automated Code Refinement Loops - Iteratively modifies source code based on performance feedback to achieve execution speedups.
- Error-Correction Feedback Loops - Injects runtime error messages and system logs back into prompts for autonomous bug fixing.
- Population Diversity Maintenance - Uses a niche-based architecture to preserve candidate solution diversity and prevent premature convergence.
- Closed-Loop Code Iteration - Provides autonomous cycles of code execution and refinement to achieve hardware-specific performance breakthroughs.
- AI Model Orchestration - Orchestrates multiple AI model providers using weighted logic to improve response accuracy.
- Automated Prompt Optimization - Programmatically refines system instructions and model prompts based on iterative performance metrics.
- Genetic Prompt Evolution - Treats system instructions as evolvable code, applying genetic algorithms to iteratively optimize agent performance.
- Deterministic Training Seeds - Applies consistent random seeds across all stochastic components to ensure scientific reproducibility.
- Weighted & Fallback Ensembles - Combines outputs from multiple LLM APIs using weighted logic and fallback strategies to increase reliability.
- Prompt Engineering - Designs and refines system instructions as evolvable code to optimize LLM performance.
- Reproducible Run Configurations - Ensures exact reproduction of experimental runs by capturing configurations and synchronizing random seeds.
- LLM Plugin Orchestrators - Implements a system to route tasks across multiple LLM APIs using weighted logic and fallback strategies.
- Experimental Reproducibility - Ensures experimental results can be repeated exactly by maintaining consistent seeds across random components.
- Code Refinement - Open-source evolutionary coding agent.
Star history
How this analysis was created: This summary and feature list were written by an AI model that read the project's README and public documentation pages. Each feature links to the documentation it came from; stars, license and language come straight from the GitHub API. The model does not read the source code, and the analysis is refreshed when the project is re-analysed. Learn more on our About page.
