Found first: a primary source the press has not covered yet.
A paper from researchers at MirroS, Tsinghua University, and Peking University identifies a linear subspace inside multimodal large language model activations that encodes continuous object coordinates across horizontal, vertical, and distance dimensions. The paper also identifies where the representations break down: viewpoint transformation is the one spatial task these internal encodings cannot resolve without external intervention.
What the source says
Zhiyuan Shi, Xiang Li, Mingsheng Long, and colleagues name the subspace S-Space and locate it by differentiating spatial-answer scores with respect to object-token activations. Causal validation confirms selectivity: editing S-Space coordinates at shift magnitude α=5 drops spatial answer retention to 62.5%, while random edits to unrelated dimensions leave answer retention at 99.8%. S-Space activates even when the model is answering non-spatial questions, indicating it is intrinsic to the representations rather than a task-specific artifact. On SpinBench, which requires perspective-taking across viewpoint shifts, Molmo2-ER's direct generation scores 60.96%; augmenting generation with S-Space rotation raises it to 94.52%.
Why it matters
MLLM spatial reasoning research has focused almost entirely on benchmarks. This paper locates the mechanism: a causally active, linearly structured coordinate map inside the model, present without prompting and encoding real 3D positions. The bottleneck is not that models lack spatial information, it is that they cannot apply coordinate transformations to that information when the viewpoint changes. For teams building spatially aware vision-language systems, the representations are already there; the problem is in the transformation step, and the S-Space framing gives a concrete place to intervene.