When to Rethink: Learning Multi-Perspective Self-Verification for Vision-Language Models

When to Rethink. Learning Multi-Perspective Self-Verification for Vision-Language Models. By Ziquan Zhu and twelve others, from the University of Leicester, Nanjing University of Posts and Telecommunications, and seven other institutions. Published on the fourth of October. Its code repository is public and real: about twenty source files.

Vision-language models reason well across modalities, yet they still produce answers that sound plausible and are wrong. Self-verification is a practical response: the model checks its own answer, with no external judge. Existing methods, though, usually lean on a single verification criterion or a fixed prompt, and the reliability estimates they give are incomplete and unstable. A check that looks at an answer from only one direction can miss what is wrong with it, and a check whose verdict changes with the wording of the prompt is hard to trust either way.

Before proposing anything, the authors analyse what drives verification quality. Two things come out. Stronger verifiers give more reliable judgments. And verification is highly sensitive to the prompt, with no single prompt consistently winning across tasks. If no one prompt is best, relying on one prompt is the weak point.

MOTIVE follows from those findings. The name stands for Multi-View Self-Verification with Reliability-Guided Selective Rethinking. Instead of judging a candidate answer from one angle, it evaluates each answer from several complementary verification perspectives, and it learns a reliability score that is aligned with whether the answer is actually correct. The goal is stated plainly in the title: knowing when to rethink. Every answer gets checked, but not every answer gets reworked.

That score is learned through what the authors call correctness-grounded multi-view verification learning. At inference it drives a single decision: accept or rethink. An answer the score marks as reliable is returned directly. An answer it marks as uncertain triggers rethinking, and that rethinking is guided by history. The model does not reconsider everything; it spends extra reasoning only where its own verification is unsure.

Experiments span diverse multimodal benchmarks and several VLM backbones, and MOTIVE consistently outperforms strong baselines of two kinds, self-verification and self-correction. Further results connect the pieces: more reliable verification leads to better accept-or-rethink decisions, and it cuts the number of unnecessary reasoning turns. The two claims depend on each other. A score that could not tell reliable answers from uncertain ones would either send everything back for another pass, or let wrong answers through untouched.

Two gains travel together here. Reliability, because answers are checked from several perspectives rather than one brittle prompt. And efficiency, because confident answers skip the rethinking step altogether. Both come without calling an outside judge, which was the point of self-verification in the first place.