Anvil links spatial learning across related tasks
In plain words
Spatial intelligence means more than recognizing objects. A model also needs to locate them, understand how they relate, and judge actions using that information. Anvil is a framework that creates and checks examples for these tasks, trains models on them, and tests their abilities.
Anvil links examples to the same scenes and objects, so training can connect one stage of reasoning to the next. Across seven public spatial tests, Anvil’s generalist model has a higher average score than its base model, though it does not score higher on every test.
The research
If a model locates an object correctly, can it use that answer to reason about the object’s relation to something else or judge a possible action? Spatial training examples often use incompatible formats, and success on one task may not transfer.
The paper asks whether shared scene-linked examples and specialist guidance can help connect these abilities.
Anvil links spatial data generation, checking, training, and assessment so a generalist can learn across related tasks with specialist guidance.
The central question is whether models can learn spatial abilities that carry across different output formats and connected stages, rather than succeed only on isolated tasks.
Anvil links records to shared scenes and entities, checks their consistency and evidence, then uses specialist guidance to train a generalist.
Across seven public spatial tests, the final Anvil-MOPD model averages 74.27%, versus 64.98% for Qwen3.5-9B. This mean supports a higher average on those tests, while a linked-task panel shows a large gap over separate-stage training. The gains are uneven, and the tests and controlled panels measure distinct things.
The seven-test mean is descriptive because published scores do not all use one evaluation protocol, and Anvil-MOPD regresses on RefSpatial.
The controlled studies use separate, task-specific panels; they support conclusions only for those comparisons, not for every spatial task or unfamiliar environment.
Analysis
Deep
- Original paper authors
- Weilin Ruan, Bohan Zhou, Jinghao Li, Yangkai Wei, Wenqian Li
- Submitted to arXiv
- License
- CC BY 4.0