Shanghai AI Laboratory open-sources SpatialASM framework for spatial reasoning in LVLMs
Shanghai AI Laboratory's InternLM team released the SpatialASM repository on GitHub, introducing a framework that internalizes Active Spatial Mentalization into large vision-language models. The repository currently contains only a README; paper, models, data, and code are marked as coming soon.
China context
- Original name
- 上海人工智能实验室
- Outside China
- Open weights · github.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can access the repository and, once released, the code and model weights, but the license is research-only, so commercial integration requires a separate license.
- For investors
- The release signals continued open-source activity from Shanghai AI Laboratory in the vision-language model space, but the non-commercial license limits direct monetization opportunities.
SpatialASM is a scalable framework that internalizes Active Spatial Mentalization (ASM) into large vision-language models (LVLMs). It consists of three components: an allocentric novel-view generator, ASM trajectory synthesis, and view-imagination reinforcement learning. The framework aims to enhance spatial reasoning by enabling models to plan informative viewpoints and imagine missing spatial evidence. The repository is licensed for research use only, with the SpatialASM-NVG model following the FLUX-NON-COMMERCIAL license.
The framework uses a two-stage training process on LVLMs, converting answer-only spatial QA data into explicit plan-imagine-reason trajectories and automatically verifiable perspective-taking tasks. A novel-view generator acts as temporary training scaffolding, so the final model does not require a world model or renderer at inference time. The allocentric viewpoint interface reaches a target perspective in one query, and Geometry Perturbation (GeoP) filtering removes unreliable synthesized views.
Developers working on spatial reasoning tasks gain a new open-source framework that could reduce the need for expensive spatial Chain-of-Thought supervision by generating it automatically from existing QA data. The research-only license limits commercial use, so enterprises must wait for a commercial license or build their own implementation.
For research teams, SpatialASM offers a method to improve spatial reasoning in LVLMs without inference-time dependencies on renderers. However, the non-commercial license restricts immediate business adoption; companies should monitor for a commercial license or consider the approach for internal research.
The repository states that the paper, models, data, and code are coming soon. A verifiable next signal is the release of the SpatialASM-128k dataset and trained model weights on Hugging Face, as well as the arXiv paper (2610.13832).