SpatialASM: Shanghai AI Lab trains LVLMs to actively mentalize viewpoints for spatial reasoning
Shanghai-AI-Laboratory · hf · 2026-10-13
Shanghai AI Lab with CUHK/Fudan introduces SpatialASM, a framework instilling Active Spatial Mentalization (ASM) into LVLMs via training-time spatial simulation. An allocentric novel-view generator grounds hypothetical viewpoints in visual evidence, producing explicit ASM trajectories and auto-constructed perspective-taking tasks. Across seven spatial benchmarks and four LVLM backbones it raises average accuracy while preserving general vision, offering a scalable path around scarce spatial CoT supervision.
More from Multimodal
- MiniCPM-V-4.7-1B, a lightweight on-device multimodal model, trends on Hugging Face — openbmb · 2026-10-13
- Eerie black-eyed AI portrait goes viral for its unsettling gaze — PierceLilholt · 2026-10-13
- AI recreates Metal Gear Solid PS1 cutscenes starring virtual idol Psyopqueen — tetsuoai · 2026-10-13
- Indie anime model MageTrail V0.4 shows artist-style learning from ~$20; 6k-artist run priced at $1-2k — Turbulent-Bass-649 · 2026-10-13
- Entire movie rendered as one sequence: FilmCraft shows cinematic AI long take — Super_Range45 · 2026-10-13
- Meta's Muse Wonder (Max) tested on one-shot 3D scenes and front-end vs top models — arena · 2026-10-13