S-Agent 2.0 adds spatial tools to general coding agents
Original title: S-Agent 2.0: Spatial Reasoning with a General Coding Agent
With Qwen3.8-27B it scores 66.5 on MMSI-Bench and 82.3 on ReVSI, above SpatialClaw.
S-Agent 2.0 equips an off-the-shelf coding agent with a command-line 3D perception toolkit and a persistent scene memory so readings are interpreted in plain spatial language. The authors report tool gains of 13.9 and 20.5 points for Qwen3.8-27B on MMSI-Bench and ReVSI, and 11.2 and 9.6 points for Gemini-3.8-Flash. With Gemini-3.8-Flash the scores are 74.1 and 90.5, versus 67.0 and 79.7 for SpatialClaw on that backbone. The system also runs on Codex and Claude Code harnesses.
- Open-backbone scores are 66.5 on MMSI-Bench and 82.3 on ReVSI.
- SpatialClaw on the same open backbone scores 62.6 and 74.5.
- Gemini-3.8-Flash reaches 74.1 and 90.5 with the toolkit.
- SpatialClaw's ReVSI tool gain falls from 13.2 to 2.7 as backbones grow.
- S-Agent 2.0's ReVSI gain on Gemini stays at 9.6 points.
- Products
- CodexClaude Code
From the original article
1 min read at huggingface.co
Spatial reasoning asks a model to work out where objects are in 3D space, how they relate to one another, and how they move. Agentic VLMs have recently brought large improvements on such tasks by letting the model actively gather evidence with 3D perception tools instead of answering in a single pass.
By Rui Gong, Ruoyu Wang, Yalun Dai · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.