Training VLMs to absorb tools eliminates dependency

SpatialCLI demonstrates that a vision‑language model can keep 73.8 % accuracy on the MindCube...