S-Space paper locates MLLMs' viewpoint-reasoning bottleneck

A new paper identifies "S-Space," a linear representation subspace inside multimodal language models that encodes object positions along horizontal, vertical, and distance axes. Causal interventions in that subspace alter spatial judgments while leaving non-spatial responses largely intact, the authors report in a preprint posted to arXiv on October 9.

The authors characterize S-Space as an "internal spatial workspace": perception writes object locations into it, verbal responses selectively read from it, and internal reasoning transforms it when the viewpoint changes. The subspace is "intrinsic rather than induced by specific data or prompts," according to the abstract, and remains readable even under non-spatial questions.

The sharpest claim is diagnostic. Using what they call a "perception-computation decomposition," the authors argue that "MLLMs can form capable spatial representations, yet fall short in the spatial transformations required to reason reliably across viewpoint changes." Direction-word readouts also surface a "systematic coupling between cardinal and image-relative reference frames." The abstract does not name which MLLMs were tested or report per-model numbers.