All the Facts, Still the Wrong Answer: Integration Gaps in Vision-Language Models.

arXiv:2610.33764v1 Announce Type: new Abstract: Multimodal reasoning requires models to combine multiple pieces of evidence into decisions that satisfy all relevant constraints. Evaluations centered on final-answer accuracy leave unclear whether failures arise