CodeEvolve: LLM-Driven Evolutionary Optimization with Runtime-Enriched Target Selection for Multi-Language Code Enhancement
Automated optimization framework achieves 15.22x speedup on enterprise Java hotspots
Researchers from Salesforce AI and academic institutions have released CodeEvolve, an evolutionary framework that harnesses LLMs to automatically optimize program performance and code quality. The system extends OpenEvolve by adding runtime-guided target selection via Java Flight Recorder (JFR) profiles, Monte Carlo Tree Search (MCTS) for candidate generation, automated code refinement, and language-specific evaluation pipelines for Java and Salesforce Apex.
In tests on a large enterprise Java codebase, CodeEvolve achieved an average speedup of 15.22x across seven hotspot functions, outperforming single-pass LLM optimization on five of them. For Apex, the full MCTS-augmented configuration produced an average of 19.5 valid programs out of 20, demonstrating that search, filtering, and refinement each contribute to more reliable optimization. The system maintains functional correctness through build validation, unit tests, performance checks, static analysis, and LLM-based review.
- 15.22x average speedup on seven enterprise Java hotspot functions
- Uses Java Flight Recorder (JFR) runtime profiles to automatically select optimization targets
- Full MCTS configuration yields 19.5 out of 20 valid programs on average for Apex
Why It Matters
Automated, correctness-preserving code optimization via LLMs and evolutionary search could slash manual tuning effort for large-scale enterprise systems.