S-Space: Exploring Spatial Workspace in Multimodal Language Models
The paper claims MLLMs carry a readable internal coordinate system for objects, but still break down on viewpoint changes.
The authors identify “S-Space,” a linear subspace where object-token activations encode horizontal, vertical, and distance information. They argue it is intrinsic to the model, not just a product of spatial prompts, because it remains readable under non-spatial questions. Causal interventions on this space can shift spatial judgments while mostly preserving non-spatial answers. Their diagnostic split points to the main failure: models can form useful spatial representations, but struggle with the transformations needed for reliable reasoning across changed viewpoints. HF Daily Papers' note
The authors identify “S-Space,” a linear subspace where object-token activations encode horizontal, vertical, and distance information. They argue it is intrinsic to the model, not just a product of spatial prompts, because it remains readable under non-spatial questions. Causal interventions on this space can shift spatial judgments while mostly preserving non-spatial answers. Their diagnostic split points to the main failure: models can form useful spatial representations, but struggle with the transformations needed for reliable reasoning across changed viewpoints. HF Daily Papers' note
score 5