Don’t Blink: Evidence Collapse during Multimodal Reasoning

Reasoning VLMs provide a tractable setting for studying how multimodal foundation models use and lose evidence during extended reasoning. We characterize this as evidence collapse: across the three reasoning VLMs studied here, attention to visual tokens, including manually annotated evidence regions, declines over the reasoning trajectory. This exposes a measurable reasoning-time failure mode: generated text can become increasingly self-consistent while its connection to visual evidence decays. Full-response entropy is the strongest transferable monitor we test, beating answer-span entropy and vision-based trajectory features on standard discrimination and calibration metrics, but it is text-only and cannot tell whether confidence is visually grounded. Adding vision naively does not close this gap because the relationship between visual disengagement and error is task-conditional. Low-entropy, visually disengaged predictions are hazardous on tasks requiring sustained visual reference, yet benign or inverted on symbolic tasks where early visual extraction suffices. A single global fusion rule cannot represent this sign flip, explaining why fusion degrades transfer. A targeted vision veto confirms the structure: it reduces selective risk where the interaction predicts danger and raises it where it predicts safety. Overall, evidence collapse is not merely a calibration failure but a task-conditional regularity of multimodal reasoning: whether visual disengagement predicts error depends on the structure of the task, not on the model alone. Code will be provided after acceptance.