The Dual Mechanisms of Spatial Reasoning in Vision-Language Models
This work clarifies the internal mechanisms of spatial reasoning in VLMs, which is incremental but important for improving multimodal tasks like image captioning and VQA.
The study investigated how vision-language models compute spatial associations, revealing that spatial information primarily originates from the vision encoder's global representations, and enhancing these representations improved spatial reasoning performance on naturalistic images.
Many multimodal tasks, such as image captioning and visual question answering, require vision-language models (VLMs) to associate objects with their properties and spatial relations. Yet it remains unclear where and how such associations are computed within VLMs. In this work, we show that VLMs rely on two concurrent mechanisms to represent such associations. In the language model backbone, intermediate layers represent content-independent spatial relations on top of visual tokens corresponding to objects. However, this mechanism plays only a secondary role in shaping model predictions. Instead, the dominant source of spatial information originates in the vision encoder, whose representations encode the layout of objects and are directly exploited by the language model backbone. Notably, this spatial signal is distributed globally across visual tokens, extending beyond object regions into surrounding background areas. We show that enhancing these vision-derived spatial representations globally across all image tokens improves spatial reasoning performance on naturalistic images. Together, our results clarify how spatial association is computed within VLMs and highlight the central role of vision encoders in enabling spatial reasoning.