"Multimodal language models can recognize visible social scenes while remaining weak at reasoning about relationships, unobservable states and evolving social dynamics."
SIV-Bench evaluates social-scene understanding, social-state reasoning and social-dynamics prediction across 2,792 video clips. The researchers find that leading multimodal models perform better on scene recognition than on deeper relational inference, providing empirical support for the chapter's emphasis on hidden social and cultural context.
View Source