algorithmicsuperintelligence/openevolve
Openevolve
OpenEvolve is an open-source framework for evolutionary computation that uses language models to drive automated optimization across multiple domains. It can evolve system prompts for large language models, refine source code across programming languages, search for optimal GPU kernel configurations, discover interpretable mathematical expressions from data, and maintain diverse populations of high-performing solutions. The framework integrates multiple evolutionary strategies, including MAP-Elites diversity mapping and island-based topologies, to avoid premature convergence and preserve a wide range of candidate variants.
The system supports checkpoint-and-resume state persistence, enabling long runs to be paused and continued without losing progress. Code regions can be marked with start and end comments so that mutation operators target only the designated sections during evolution. Evolution runs can be configured through a central YAML file, with deterministic seeding for reproducibility across different machines. The framework combines multiple language model providers using an intelligent ensemble with weighted averaging and fallback strategies, and it applies multi-stage evaluation cascades to filter candidates through quick preliminary tests before expensive comprehensive evaluation.
OpenEvolve provides an interactive web-based visualization that renders live evolution trees, performance charts, code diffs, and a MAP-Elites grid for real-time tracking. It collects execution artifacts and error feedback from previous generations to inject into subsequent prompts, improving the evolutionary guidance. The project is configured via a single YAML file and can be installed for immediate use.
Features
- Automated Prompt Engineering - Applying evolutionary search to discover and refine system prompts for large language models, optimizing them for accuracy and other criteria.
- Genetic Prompt Evolution - Evo evolves system prompts or code automatically to improve LLM performance on benchmark tasks using evolutionary search.
- Evolutionary Code Optimizers - Using evolutionary algorithms to automatically mutate and select source code variants that improve performance or correctness across multiple programming languages.
- Evolutionary GPU Kernel Tuners - Automatically searching for optimal GPU kernel implementations for neural network operations using evolutionary methods to maximize performance.
- Evolutionary Algorithm Discovery - Evo transforms a basic random search into advanced algorithms like simulated annealing through iterative code mutation and performance evaluation.
- Fitness Functions - Evo defines a fitness function that returns a numeric score and optional artifacts to measure solution performance on a given task.
- Weighted & Fallback Ensembles - Combines multiple language models using weighted averaging and fallback strategies to improve robustness.
- Multi-Provider Abstractions - Evo uses a universal API to combine multiple AI models with intelligent ensemble and fallback strategies for robustness.
- Prompt Quality Scoring - Evo combines task accuracy with qualitative LLM evaluation on clarity, specificity, robustness, and format to rank prompts.
- Quality-Diversity Optimization - Implements quality-diversity optimization using MAP-Elites and island topologies to maintain diverse solution populations.
- Tool Error Feedback - Evo captures artifacts and error messages from previous generations and injects them into subsequent prompts to improve evolution.
- Island-Based Topologies - Separates populations into isolated groups that periodically exchange individuals to avoid premature convergence.
- Quality-Diversity Grids - Maps candidate solutions onto a multi-dimensional grid to maintain a diverse archive of high-performing variants.
- Cascading Evaluation Filters - Filters candidates through a quick preliminary test before a more expensive comprehensive evaluation.
- Symbolic Expression Evolution - Evo modifies code within marked blocks iteratively to discover mathematical expressions that minimize error against training data.
- Symbolic Regression - Discovers mathematical expressions by iteratively mutating and selecting symbolic representations.
- Symbolic Regression Engines - Evolves mathematical expressions from raw data using genetic programming and execution feedback for symbolic regression.
- Artifact Logging - Evo stores intermediate logs, results, and debugging information during evaluation for detailed analysis of evolved solutions.
- Island-Based Topologies - Evo configures preset templates for island-based evolutionary algorithms with sensible defaults and scenario-specific tuning options.
- Evolutionary Run Checkpoints - Serialises evolution state to disk so runs can be paused and resumed without losing progress.
- Real-Time Evolution Visualizations - Evo displays an interactive web interface with evolution trees, performance charts, code diffs, and a MAP-Elites grid for live tracking.
- GPU Kernel Evolution - Evo searches automatically for optimal GPU kernel configurations using evolutionary algorithms to improve neural network operation performance.
- Evolutionary Run Dashboards - Renders live evolution trees, performance charts, code diffs, and diversity grids in a browser interface.
- Stateful Run Resumption - Evo launches evolution runs from scratch or resumes from saved checkpoints, retaining all state for seamless continuation.
- GPU Kernel Evolution - An automatic optimizer that uses evolutionary algorithms to find the fastest GPU kernel implementations for neural network operations.
- Evolutionary Tree Visualizations - Evo renders a graphical tree showing the ancestry and performance of generated code variants across generations.
- Evolvable Code Region Markers - Marks delimited code blocks with start and end comments so mutation operators target only those sections.
- Analytical Reproducibility - Evo seeds every component deterministically so that the same evolution can be exactly reproduced across different machines.
- Evolving - Evo evolves sorting algorithms that dynamically adjust behavior based on input data characteristics to improve performance.
- Competitive Programming Solution Evolution - Evo evolves a program that passes all test cases on an online judge by iteratively mutating and selecting candidate solutions.
- Cascade Prompt Evaluations - Evo tests prompts in two stages (quick 10-sample then comprehensive 40-sample) to efficiently filter low-quality candidates.
