Prayag Tiwari

From Perception to Integration: Revisiting the Internal Dynamics of Reasoning in Vision-Language Models

Vision-language models (VLMs) can answer simple visual questions, but often struggle when one question requires several visual judgments. We study this gap with controlled tasks for feature binding, numerosity, spatial relations, and amodal completion, together with a Composite task that combines them. Matched counterf...