Verify a patch by reconstructing the problem backwards from it — without showing the verifier the original issue
RETRACE (arXiv 2608.08950, Aug 9) fixes a blind spot in agent self-review: existing verifiers read the patch under the same interpretation that produced it, so they inherit the original misreading. Its three stages are forward reconstruction (build an explicit repair rationale from the issue plus the agent's decision trajectory), backward reconstruction (working only from the patch and trajectory, with the original issue withheld, infer what problem this patch appears to solve, then compare that inference against the real issue), and reconciliation (diagnose any misalignment and either approve or issue targeted revision guidance). Training-free, it lifted Pass@1 by 7.0% and 3.6% on mini-SWE-agent over SWE-bench Verified and produced comparable gains on OpenHands with no modification — the information-hiding trick is the reusable idea for any LLM-judge setup.
↳ Follow the thread