The authors built a system where a small VLM (Qwen3.5-4B) uses external tools for 3D object detection, metric depth, and deterministic geometric calculations to answer questions about distances, sizes, and directions from video frames. Tools lift camera-space detections into a shared world frame. On the ReVSI-Bench benchmark, the tool-augmented model achieved large gains on absolute distance (0.46 to 0.74 MRA), relative distance (39.1% to 67.4%), and relative direction (25.9% to 73.4%) compared to the no-tool baseline. Improvement on object size was marginal (0.58 to 0.61). By swapping real detectors for ground-truth boxes, the authors show the perception detector is the main bottleneck, with orchestration costing only 0.03 MRA. This work demonstrates that explicit geometric tools can overcome VLMs' weakness in metric 3D reasoning and provides a clean way to separate perception errors from reasoning errors.
Geometric tools let a small vision-language model measure 3D scenes
Explicit detection, coordinate projection and geometric solvers substantially improved distance and direction estimates, while detector errors continued to limit size measurement.
Academic
Technical University of Munich
Research Digest··2 min read
In one line
Tool-augmented VLMs improve metric spatial reasoning on distance and direction but not object size.
Before you read this
Three ideas this paper builds onfine-grained spatial reasoningReasoning precisely about object positions, orientations, and spatial relations across planar, depth, and temporal dimensions for robotics and navigation.72 papers →vision-language modelA model that handles tasks combining visual and textual information, such as editing images from conversation or reasoning across images and text.265 papers →multi-turn tool useThe capability of an agent to invoke external tools sequentially over multiple conversational turns to complete complex tasks.52 papers →
Read it as
Grange and Glantz equipped Qwen3.5-4B with callable tools for 3D object detection, metric depth, coordinate projection and deterministic spatial calculations.
The authors evaluated Qwen3.5-4B on four metric spatial reasoning tasks from the ARKitScenes subset of ReVSI-Bench: absolute distance, object size, relative distance and relative direction.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.