LLM builds an internal spatial map but ignores it — the unused map is more accurate (80% vs 73%)
Cristhian-AI-Math · reddit · 2026-11-13
A researcher tested whether an LLM builds spatial representations when reading text. It does — but the model answers via a shortcut and ignores its own map.
- Tested on Gemma 3 12B across 300 new scenes and 5,350 questions
- On questions requiring multi-step spatial reasoning ("Cy vs Bo"), answers read off the model's unused internal map scored 80% accuracy vs 73% for the model's actual answers
- On directly stated questions the two were nearly identical (89% vs 91%, 90% vs 90%)
The finding suggests spatial reasoning could improve if models were guided to exploit their internal spatial representations.
Related event: Gemma 3 12B Builds Internal Spatial Maps but Ignores Them(2 posts)→
More from Research
- Ropedia Xperience 2: 1,589 3D scans and 131 hours of egocentric data for physical AI — ropedia-ai · 2026-11-13
- Zhejiang U & Baidu's Agent-R²: 4B agent beats larger open-source baselines on GDPval — ZhejiangUniversity · 2026-11-13
- New model shows row norms ≈1 across all layers, echoing nGPT's hypersphere hypothesis — ChengleiSi · 2026-11-13
- Agoric Open Systems Papers: overlooked 1988 foundation for AI safety, says researcher — sebkrier · 2026-11-13
- Qwen3-32B Has a Body Schema: Wheels Map to Heels, Grippers to Palms — Sauers_ · 2026-11-13
- scaling01 replicates looping transformer: 4B model reasoning climbs 42.9→52.4 over 4 loops — scaling01 · 2026-11-13