This technology is a context-aware refinement module that transforms partially correct patches generated by large language models (LLM) into correct fixes, improving the accuracy of existing automatic program repair systems.
State-of-the-art LLM-based automated program repair (APR) tools generate code patches by reasoning over issue descriptions, codebases, and regression test suites. However, these tools frequently produce partially correct fixes that address individual instances of a bug while failing to resolve related cases or the underlying defect. As a result, many patches still require manual review and refinement before they can be safely merged, reducing the efficient gains of automated repair. Current APR systems lack systematic methods for refining these near-correct patches into complete, reliable fixes.
This technology is a software refinement module that transforms partially correct code patches generated by existing automatic program repair (APR) tools into complete and correct fixes. The module analyzes the issue description, draft patch, and surrounding code context to generate, evaluate, and integrate improved code edits before validating the final patch against the project’s test suite. Because the module refines the output of existing repair tools rather than replacing them, it can be integrated into both autonomous-agent and fixed-workflow APR systems.
This technology was evaluated on the SWE-Bench Lite and SWE-Bench Verified benchmarks of real-world GitHub issues. When integrated with five automated program repair tools, it improved bug-resolution rates by an average of approximately 14% and resolved several issues that none of the baseline tools were able to fix.
Patent Pending