Researchers propose SpatialClaw, a training-free framework that adopts code as the action interface to improve open-ended 3D and 4D spatial reasoning in vision-language models. Unlike existing agents that rely on single-pass execution or rigid tool calls, SpatialClaw maintains a stateful Python kernel pre-loaded with input frames and perception primitives.

  • The agent writes one executable cell per step conditioned on all prior outputs, allowing flexible composition of perception results.
  • Evaluated across 20 spatial reasoning benchmarks, SpatialClaw achieves 59.9% average accuracy.
  • This performance outperforms the recent spatial agent by +11.2 points with consistent gains across six VLM backbones from two model families.

The approach enables agents to adapt their analysis to intermediate text and visual observations without benchmark- or model-specific adaptation.