S-Space: Researchers Uncover a Linear Spatial Workspace Inside Multimodal LLMs
MirroS-Lab · hf · 2026-10-12
A MirroS technical report uncovers S-Space, a linear representation subspace in multimodal LLMs where object tokens encode continuous coordinates along horizontal, vertical, and distance axes.
- S-Space is intrinsic rather than prompt-induced and stays readable under non-spatial questions.
- Perception writes object locations into it, internal reasoning transforms it for viewpoint changes, and verbal output accesses it selectively—causal interventions alter spatial judgments while preserving non-spatial responses.
- As a diagnostic tool, it reveals a key bottleneck: MLLMs form capable spatial representations but fail at the transformations needed for cross-viewpoint reasoning.
More from Research
- METR's AI time-horizon chart hasn't updated since May, leaving frontier model gains unmeasured — haider1 · 2026-10-12
- 15 detectors, 3 benchmarks: best prompt-injection detector catches 2% of real agent injections — lulzxdxdxd · 2026-10-12
- HANDBOOK.md benchmark: top agents pass just 36.2% of 65 policy-heavy agentic tasks — bibryam · 2026-10-12
- AI finds solutions hiding in plain sight across 22 scientific fields, Science reports — Spare-Dingo-531 · 2026-10-12
- Terence Tao blog: what mathematicians should know about Lean's reliability and AI — thomasahle · 2026-10-12
- Michael Levin publishes a starter pack for his lab's work on diverse intelligence — drmichaellevin · 2026-10-12