RETRACE: verify the patch by reconstructing the problem it actually solves

Chenglin Li, Yisen Xu, Zehao Wang, Shin Hwei Tan, and Tse-Hsun (Peter) Chen (Concordia) attack the post-generation blind spot: once a coding agent emits a patch, nothing independently asks whether that patch solves the reported issue rather than a neighboring one. Self-refine reviews under the same interpretation that produced the patch; sample-and-select broadens candidates without a per-patch alignment signal. RETRACE is a training-free wrapper around an existing scaffold. Forward reconstruction builds an explicit repair rationale (evidence, reasoning chain, intended repair) from the issue and trajectory without seeing the patch. Backward reconstruction withholds the original issue and infers, from patch + trajectory alone, the problem the implementation appears to address. An alignment verdict (Same / Partial / Different) plus structured disagreements feeds reconciliation, which attributes mismatch to reasoning, implementation, or both and either submits or returns a targeted revision plan. The agent’s tools stay unchanged; RETRACE only gates submit vs revise.

On SWE-bench Verified (500): mini-SWE-agent Pass@1 rises 56.2% → 63.2% (GPT-5-mini, +7.0) and 75.8% → 79.4% (MiniMax-M2.5, +3.6). Self-Refine with comparable extra inference drops both baselines. On a 120-issue OpenHands transfer slice, gains remain large (37.5% → 56.7% / 62.5% → 70.0%) without scaffold edits. Ablations: forward-only and backward-only each lift, rescue different failure sets, and full reconciliation adds more. Prompt-cache-heavy stages keep per-issue cost at or below baseline.

Why it matters for code review

Continuous verification under volume needs a signal that is not the writer’s self-story and not only green unit tests the agent itself wrote. Review always-on merge lanes for RETRACE-shaped independence: reconstruct the problem the patch implies without re-feeding the ticket, then reconcile before human or auto-merge. Overnight harnessed momentum compounds when each candidate faces that check; bursty Claude mastery that self-critiques in the same context still ships interpretation errors. Pair with ECLoop’s pre-commit evidence gate and Zhang VeriHarness’s disagreement/consensus environmental checks when the factory must verify both when to edit and whether the finished patch still matches the ticket.