A new arXiv preprint introduces S-Agent 2.0, a general coding agent aimed at spatial reasoning. The paper defines spatial reasoning as the task of working out where objects are in 3D space, how they relate to one another, and how they move.
The authors note that agentic vision-language models (VLMs) have recently brought large improvements on such tasks by letting the model actively gather information. The abstract is truncated in the source, so the exact mechanism for gathering information is not specified, but the framing suggests the agent uses coding or tool use to collect what it needs.
Because only the abstract is available, details of the architecture and evaluation are not yet clear. The apparent contribution is in applying a general coding agent to spatial reasoning rather than building a task-specific model, which could make the approach more flexible across different 3D reasoning problems.