S-Agent gets 46.4% on MMSI-Bench by turning spatial reasoning into action chains

机器之心 · wechat · 2026-07-24

South China University of Technology’s S-Lab and Ropedia propose S-Agent, a spatial reasoning framework that turns spatial understanding from one-shot answering into an action chain.

A VLM reads the question, plans the next step, and delegates geometry-heavy work to specialized models such as DA3.
The system uses a three-level tool stack: 2D perception, 3D alignment, and spatial experts that convert raw geometry into usable evidence.
The paper reports 46.4% on MMSI-Bench zero-shot and 60.0% on ViewSpatial-Bench; with S-300K trajectory distillation, an 8B model reaches 41.6 / 46.8 on MMSI/ViewSpatial.
The authors argue that spatial intelligence should be evaluated as a traceable, reusable agent loop that can accumulate evidence, not as a single answer guessed from a video.

Original post →

S-Agent gets 46.4% on MMSI-Bench by turning spatial reasoning into action chains

More from Multimodal