Multimodal language models are maddeningly inconsistent at spatial reasoning: ask whether one object sits to the left of another and the answer may land right or wrong with no obvious pattern. A paper posted to arXiv on October 9 by the MirroS Lab team offers a counterintuitive diagnosis: these models do have spatial sense — they maintain an internal three-dimensional coordinate system — and the real weakness lies elsewhere (arXiv:2610.13527). Within three days the paper climbed to #2 on the Hugging Face Daily Papers board with 199 upvotes.
A 3-axis coordinate system hidden in intermediate layers
The researchers isolated a linear subspace in the models' intermediate activations, naming it S-Space: object-token representations encode continuous coordinates along horizontal, vertical, and distance axes — not discrete left/right labels but numeric positions that can be read out. The discovery method is refreshingly practical: differentiate spatial-answer scores with respect to object-token activations and keep the directions that matter most. Crucially, the coordinates are intrinsic — they remain readable under non-spatial questions, so they are not an artifact of particular prompts.
A spatial workspace you can intervene on
The paper frames this subspace as an internal spatial workspace, with three layers of evidence. Perception writes to it: object locations are encoded as spatial representations that directly support accurate spatial judgments. Reasoning transforms it: when the viewpoint changes, internal reasoning operates on the representation. Verbal responses access it selectively: causal interventions that rewrite S-Space coordinates change spatial judgments while largely preserving non-spatial answers — the hardest evidence that this is a genuine workspace. Case studies show it assigns locations to unseen entities and to the observer, and direction-word readouts expose systematic coupling between cardinal and image-relative reference frames.
The bottleneck is transformation, not perception
The most diagnostic result is a perception-computation decomposition: multimodal models form capable spatial representations yet fall short on the spatial transformations that reliable viewpoint-change reasoning demands. The repository README states it plainly — explicit computation over S-Space can outperform native reasoning. In other words, the models are not missing spatial sense; they fail to invoke it reliably.
Open source and reproducible
The code is on GitHub under Apache 2.0 with 247 stars, and pretrained S-Space axes ship with the repository. Experiments cover Molmo2-ER, MolmoAct2, Qwen3.5-4B and Qwen3.6-27B, with benchmarks including COCO, CV-Bench, EmbSpatial, SpinBench and HStar; reproduction is a matter of running the bundled scripts.
Beyond being an interpretability finding, the paper turns vague spatial failures into two decidable failure modes: was the representation never built, or was the transformation never performed? For embodied-AI and robotics teams — whose agents live and die by viewpoint changes — that distinction is worth money. And for model builders the implication stings: rather than hoping chain-of-thought spontaneously learns spatial transformation, wire explicit spatial computation into the architecture.