SpatialCLI: Learning to Reason With Spatial Tools, Then Without Them

SpatialCLI is a three-stage framework that teaches vision-language models to use spatial perception tools and then internalize those capabilities.

The team also released SpatialCLI-Bench, a 516-example benchmark covering localization, segmentation, depth, and pose.

Applied to Qwen3-VL-8B-Instruct, it delivered large gains on MindCube and retained more performance after tools were removed.

With tools enabled, it even outperformed GPT-5.6 Sol on MindCube, highlighting a path to stronger embodied visual reasoning with less deployment-time tool dependence.