The abstract for the Anvil paper argues that embodied spatial intelligence demands more than visual recognition. Models must localize entities, reason about their relations, and express that knowledge in forms that can directly guide action. This goes beyond standard perception tasks.

The authors identify a key challenge: joint training can conflate these distinct knowledge types. When a model learns everything at once, it may mix up localization, relational reasoning, and action-oriented output. Anvil addresses this through a method called knowledge recirculation, which is presented as a way to scale these capabilities.

The available abstract is brief, so the exact mechanics of knowledge recirculation are not described in the text. Readers are directed to the full arXiv manuscript for details. This summary reflects only the stated motivation and proposed approach from the abstract.