A perception tool can return the right measurements and still leave a model unable to answer the question. The central insight here is that spatial reasoning needs a bridge from raw evidence to usable conclusions: retain scene evidence across steps, then express each reading in language the agent can reason with.
Key Findings
- The command-line toolkit combines 3D perception with persistent scene memory and plain-language interpretations, keeping conclusions grounded in geometric evidence.
- On MMSI-Bench and ReVSI, adding the toolkit improved Qwen3.8-27B by 13.9 and 20.5 points, and Gemini-3.8-Flash by 11.2 and 9.6 points, compared with the same agents without tools.
- With Qwen3.8-27B, the system scored 66.5 and 82.3 on the two benchmarks, ahead of SpatialClaw on the same backbone at 62.6 and 74.5. With Gemini-3.8-Flash, it scored 74.1 and 90.5, compared with SpatialClaw’s 67.0 and 79.7.
- The toolkit works with general coding-agent harnesses including Codex and Claude Code, suggesting the approach is not tied to one agent interface.
Method and Positioning
Rather than relying on a single visual impression, the agent gathers evidence through perception tools, preserves relevant object-centered observations in scene memory, and interprets readings in spatial language. This targets a practical failure mode in tool-using vision-language systems: evidence acquisition may succeed even when evidence integration does not.
Who It Suits—and the Trade-off
A useful read for researchers building agentic vision-language systems or evaluating persistent 3D reasoning. It is less directly useful if you need a ready-to-deploy spatial assistant: the contribution is an agent-and-toolkit approach evaluated on benchmarks, and its gains depend on the underlying model and harness.