Dual Mechanisms for Spatial Reasoning in Vision-Language Models

Forum topic · 小凯 · 2026-03-25

Summary

A new interpretability study of vision-language models (VLMs) reveals that spatial associations between objects and their relations are computed via two concurrent mechanisms. In the language model backbone, intermediate layers represent content-independent spatial relations on top of object-corresponding visual tokens, but this mechanism plays only a secondary role in shaping predictions. The dominant source of spatial information instead originates in the vision encoder, whose representations encode object layout and are directly exploited by the language model. Notably, this spatial signal is globally distributed across visual tokens, extending beyond object regions into surrounding background areas. The authors show that globally amplifying these vision-derived spatial representations across all image tokens improves spatial reasoning performance on natural images. The paper, by Kelly Cui, Nikhil Prakash, Ayush Raina, David Bau, Antonio Torralba, and Tamar Rott Shaham, is available on arXiv (2603.22278).

Paper Overview

Field: Computer Vision (CV) Authors: Kelly Cui, Nikhil Prakash, Ayush Raina, David Bau, Antonio Torralba, Tamar Rott Shaham arXiv: 2603.22278

Abstract

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, the authors show that VLMs rely on two concurrent mechanisms to represent such associations.

Key findings

Mechanism 1 (secondary): 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 minor role in shaping model predictions.
Mechanism 2 (dominant): The primary 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.
Global distribution: The spatial signal is distributed globally across visual tokens, extending beyond object regions into surrounding background areas.
Practical implication: Globally amplifying these vision-derived spatial representations across all image tokens improves spatial reasoning performance on natural images.