CodeEvolve: LLM-Driven Code Optimization Framework
CodeEvolve is a cutting-edge system that uses Large Language Models (LLMs) to improve the quality of code and overall program efficiency. It evolves from OpenEvolve by adding features like runtime-guided target selection, Monte Carlo Tree Search (MCTS), and automated code refinement specifically for Java and Salesforce Apex. By utilizing Java Flight Recorder (JFR) profiles, it creates weighted graphs to pinpoint optimization opportunities that heavily impact execution costs, thereby reducing the need for manual bottleneck identification. For each target, CodeEvolve suggests edits, which undergo thorough validation through build checks, unit tests, performance assessments, static analysis, and LLM reviews, ensuring that only effective and correct changes are implemented. In real-world applications, it significantly boosts performance and code metrics while maintaining correctness.
Key facts
- CodeEvolve is an evolutionary framework for code optimization using LLMs.
- It extends OpenEvolve with runtime-guided target selection.
- Uses Monte Carlo Tree Search (MCTS) for automated code refinement.
- Supports Java and Salesforce Apex with language-specific evaluation pipelines.
- Employs Java Flight Recorder (JFR) profiles to build weighted component graphs.
- Reduces reliance on manual bottleneck identification.
- Evaluates candidate edits through build validation, unit tests, performance checks, static analysis, and LLM-based review.
- Retains only variants that preserve functional correctness.
- Achieves performance improvements on a large enterprise Java codebase.
Entities
Institutions
- arXiv
- OpenEvolve