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

Vision-language models (VLMs) often struggle with precise spatial tasks like localization, depth estimation, and pose detection, despite their strong general reasoning abilities. While specialist vision models can perform these tasks accurately, they lack the high-level reasoning needed to coordinate multiple perceptual steps. This paper investigates whether a VLM can learn to use these specialist models as tools and subsequently internalize those perceptual capabilities to perform complex spatial reasoning without external assistance.

SpatialCLI is a three-stage framework designed to bridge the gap between general reasoning and specialist perception:

1. Call: The VLM is equipped with a suite of specialist vision tools (Locate, Segment, Depth, and Pose). The model learns to invoke these tools during inference to gather fine-grained visual evidence.
2. Learn: The model undergoes Cold-Start Supervised Fine-Tuning (SFT) using teacher trajectories, followed by agentic Reinforcement Learning (RL) to optimize tool selection, argument generation, and result utilization.
3. Internalize: Successful tool-use trajectories are converted into natural-language reasoning chains. The model is then trained on these chains using a dual-view approach, which forces it to learn the reasoning logic directly while preserving its ability to use tools when necessary.

The authors introduce SpatialCLI-Bench, a 516-example benchmark for compositional spatial reasoning. Experimental results show that SpatialCLI significantly boosts performance across various models. For example, the Qwen3-VL-8B-Instruct model improved from 29.3% to 84.6% on the MindCube benchmark when using tools. Crucially, after the internalization stage, the model retained a 73.8% success rate even without tool access, demonstrating that the model successfully learned to perform complex spatial reasoning natively.