acceptodds
Desk rejected at ICLR 2027

Defect Identity Determines When Verifier Feedback Should Constrain Search

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Localized verifier feedback carries a control signal: it indicates how broadly a critique system should search next. We formulate post-verifier repair as defect-conditioned search control, with defect identity serving as the routing state. Defect-Preserving Repair (DPR) restricts admissible updates to evidence-grounded edits that preserve failure family, target span, and evidence type; a boundary router renews search when the diagnosis indicates a transition. Across the reported matched-budget comparisons, DPR reaches Correction Accuracy on a -instance correction study panel, points above unrestricted post-feedback rewriting. A replay-frozen comparison retains a -point advantage, while citation-grounded QA and instruction-following study panels preserve positive utility effects. The count-derived action effect changes from points on same-defect cases to on transition-or-switch cases. A five-fold cross-fitted router on replay-frozen states uses this boundary to reach Corr-ACC, points above always applying DPR. These results establish defect identity as an actionable state variable for allocating post-verifier search.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
alignment safety and oversight, Defect Identity Determines When Verifier Feedback Should Constrain Search

Related papers

Discussion (0)

Sign in to comment.