A new arXiv preprint, "DeSpatial: The devil of Agentic Spatial Reasoning Lies in 3D Itself," examines why AI agents still struggle with spatial reasoning. The authors note that agents can call reconstruction and segmentation tools, but the outputs are not yet dependable enough to serve as reliable 3D scene information.
The problem, they argue, is that predictions are inconsistent and object identities are fragmented across different views or steps. A simple oracle experiment—described only briefly in the abstract—appears to confirm that fixing the 3D input matters more than improving the agent's reasoning loop.
This digest is based solely on the arXiv abstract, so the full methodology and results are not yet available. Still, the paper's central claim is clear: the weak link in agentic spatial reasoning may be the 3D data itself, not the agent.