When Good Verifiers Go Bad: Silent Negative Transfer in Verifier-Guided VLM Training

T1Covered by 1 source · first reported by arXiv at 24 Sept, 04:00 UTC

arXiv:2606.14629v2 Announce Type: replace-cross Abstract: Verifier reliability is not portable across tasks. A verifier-guided self-DPO pipeline with genuine held-out gains on MathVista (+9.6 points on self-training data, +8.0 held out) can be harmful on MMMU. The failure is invisible from the…

science

Sources

• T1When Good Verifiers Go Bad: Silent Negative Transfer in Verifier-Guided VLM TrainingarXiv24 Sept, 04:00