MirroS finds S-Space: a manipulable spatial workspace inside multimodal models
HeyAmit_ · x · 2026-09-09
MirroS research identifies S-Space, a structured internal spatial map in multimodal models analogous to human spatial cognition. Researchers can read, manipulate, and track it during reasoning. Crucially, changing an object's spatial coordinates inside the representation systematically flips the model's spatial judgments — left becomes right, above becomes below — suggesting S-Space plays a causal role in spatial reasoning, not just an inspectable representation. Released with an interactive blog, technical report, and all experiment code.
More from Research
- Future foundation models will be built on egocentric, calibrated multiview data — ducha_aiki · 2026-09-09
- NumeriaPlus: 300-hour egocentric activity dataset with dense ground truth for world models — ducha_aiki · 2026-09-09
- OpenAI claims 10,000 agents proved Navier-Stokes blowup in 88 hours; Clay Institute won't accept it — AdaptiveAgents · 2026-09-09
- LLM Zelda Level Generator Hits Diversity Metric but Produces Visually Boring Rooms — Amidos2006 · 2026-09-09
- Honest limitation of an evolutionary Zelda generator: high fitness ≠ visual diversity — Amidos2006 · 2026-09-09
- LLM-evolved global functions are readable: a ~5-line counting function is most common — Amidos2006 · 2026-09-09