AI/ ai · developer-tools · program-repair · benchmarks

ARISE Lifts AI Bug-Fixing Accuracy on Open-Source Repos

A new graph-based toolset called ARISE helped an open-source coding model resolve 22% of real GitHub issues, up nearly 5 points from the unmodified baseline.

An academic toolset is making AI-driven bug repair meaningfully more precise — without requiring a proprietary model.

ARISE (Agentic Repository-level Issue Solving Engine) is a framework-agnostic graph layer that maps not just how a codebase is structured, but how data flows through it. Researchers tested it by mounting it on SWE-agent, a well-known repair scaffold, using the open-source Qwen2.5-Coder-32B-Instruct model as the backbone. On SWE-bench Lite — 300 real GitHub issues spread across 11 Python repositories — that combination resolved 22.0% of issues, compared to 17.3% for the unmodified SWE-agent running the same model. The 4.7 percentage-point gain came almost entirely from sharper fault localization: the system's ability to pinpoint the right function before attempting a fix (Function Recall@1) rose from 0.43 to 0.60, a 40% relative gain, and line-level precision (Line R@1) jumped from 0.26 to 0.41, a 58% relative gain.

The localization gap matters because most repair frameworks already handle patch synthesis reasonably well once they know where to look. ARISE's core primitive — data-flow slicing, a single API call that traces which statements define or consume a given variable — gives an agent the semantic granularity that purely structural graphs lack. The researchers also ran the same toolset on a second host framework, suggesting the gains are not scaffold-specific.

SWE-bench Lite has become the de facto leaderboard for this class of research, and 22% with a fully open-source backbone is a competitive number — though proprietary-model systems routinely score higher. The authors frame ARISE as a drop-in layer for future work, which is a reasonable pitch, but whether it holds up on codebases outside Python remains an open question.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →