RETRACE: From Entangled Repair Histories to Reusable Experience for CI Repair
RETRACE turns messy CI repair histories into problem-level examples that LLM agents can reuse.
The paper says pull requests often mix several CI failures, failed fixes, reverted edits, and unrelated changes, making it hard to see which code change solved which problem. RETRACE reconstructs that alignment by looking backward from the passing revision and forward through the commit history, then using CI execution evidence to reconcile the two views. On CI-REPAIR-BENCH, it raises mini-SWE-agent Pass@1 from 19.6% to 31.9% with MiniMax-M2.5 and from 23.3% to 32.8% with DeepSeek-V4-Flash. On a matched subset, Codex improves from 15.5% to 27.5%. ArXiv · AI/CL/LG's note
The paper says pull requests often mix several CI failures, failed fixes, reverted edits, and unrelated changes, making it hard to see which code change solved which problem. RETRACE reconstructs that alignment by looking backward from the passing revision and forward through the commit history, then using CI execution evidence to reconcile the two views. On CI-REPAIR-BENCH, it raises mini-SWE-agent Pass@1 from 19.6% to 31.9% with MiniMax-M2.5 and from 23.3% to 32.8% with DeepSeek-V4-Flash. On a matched subset, Codex improves from 15.5% to 27.5%. ArXiv · AI/CL/LG's note
score 4