How MIT’s Genesis System Cuts Bug‑Fix Time in Half with Automated Patch Generation

MIT’s research team has created a system called Genesis that automatically learns from previously applied patches, generates patch templates, and evaluates candidate patches. The system’s bug‑fix success rate is almost twice that of the best hand‑crafted template systems, allowing programmers to spend roughly half the time fixing bugs.

Key Contributions

Template‑AST and generator‑based transformations that abstract patch and application specifics to capture common patterns across multiple patches.

A novel patch generalization algorithm that automatically creates transformations capturing common generation patterns from a set of patches.

An innovative search‑space reasoning algorithm, including a new sampling method and an ILP‑based selection solution, to identify promising subsets of training patches.

A complete patch‑generation system with bug localization and candidate‑patch evaluation, demonstrated on large‑scale real‑world applications.

Experimental Results

In experiments on 41 bugs from 49 test cases across 21 open‑source Java projects, Genesis correctly repaired 21 bugs, whereas the best existing hand‑written Java patch generator repaired only 11. The system selected 85 diverse templates from an initial pool of 5–10 hand‑crafted ones, improving both coverage and precision.

All experimental data, including Genesis source code, inference templates, and generated patches, are available at http://groups.csail.mit.edu/pac/patchgen/ .