Researchers find S-Space: multimodal models encode left/right/up/down/depth in a linear subspace

jiqizhixin · x · 2026-09-18

MirroS researchers report discovering S-Space (Spatial Workspace) inside multiple multimodal models: by analyzing intermediate activations, they found objects' left-right, up-down, and near-far positions are written into a stable low-dimensional representation space — a continuous linear subspace spanned by three directions, where reading along each axis yields the object's continuous coordinate. Striking examples include the model inferring an off-screen ball's position from a diving goalkeeper, and the abstract concept "socialist" landing on the left side of the internal space, mirroring the "left wing" spatial metaphor in language.

Original post →

More from Research

Research channel →