S-Space: Exploring Spatial Workspace in Multimodal Language Models
Researchers identify S-Space, a linear subspace where multimodal models encode object coordinates.
The paper identifies S-Space, a linear subspace in multimodal language models where object-token activations encode continuous horizontal, vertical, and distance coordinates. The authors say the representation is intrinsic and remains readable under non-spatial questions. Perception writes locations into the subspace, internal reasoning transforms it for viewpoint changes, and causal interventions alter spatial judgments while largely preserving non-spatial responses. A perception-computation split finds that models can form spatial representations but fail at the transformations needed for reliable viewpoint reasoning.
- S-Space encodes horizontal, vertical, and distance coordinates in object tokens.
- The subspace stays readable even when the prompt is not spatial.
- Interventions change spatial judgments while mostly preserving other answers.
- Viewpoint transformation, not perception, is the main reasoning bottleneck.
From the original article
1 min read at huggingface.co
Multimodal language models (MLLMs) exhibit promising spatial reasoning capabilities, yet the internal mechanisms supporting them remain poorly understood. We uncover S-Space, a linear representation subspace in which object-token representations encode continuous coordinates along horizontal, vertical, and distance axes.
By Zhiyuan Shi, Xiang Li, Xingzhuo Guo · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.