Researchers publishing on Hugging Face have introduced S-Space, a linear representation subspace within multimodal language models (MLLMs) in which object-token representations encode continuous coordinates along horizontal, vertical, and distance axes. The work, titled S-Space: Exploring Spatial Workspace in Multimodal Language Models, aims to shed light on the internal mechanisms behind spatial reasoning in these models.
According to the paper, the axes were identified by differentiating spatial-answer scores with respect to object-token activations. Crucially, S-Space is described as intrinsic rather than induced by specific data or prompts, remaining readable even under non-spatial questions.
The authors argue that S-Space functions as an internal spatial workspace. Perception encodes object locations into it, forming spatial representations that can be read out directly to support accurate spatial judgments. Internal reasoning then transforms this workspace for changing viewpoints, while verbal responses selectively access it — a claim supported by causal interventions that altered spatial judgments while largely preserving non-spatial responses.
Case studies further indicate that S-Space assigns locations to unseen entities and to the observer, and connects physical directions with abstract linguistic meanings, pointing to both the capabilities and limitations of large-scale language training.
Finally, the researchers use S-Space as a diagnostic tool for spatial reasoning failures. Direction-word readouts reveal systematic coupling between cardinal and image-relative reference frames. More importantly, a perception-computation decomposition exposes a key bottleneck: MLLMs can form capable spatial representations, yet fall short in the spatial transformations required to reason reliably across viewpoint changes.