Event

S-Agent 2.0 improves spatial reasoning performance with a toolkit

1 reports · 1 sources

The research team proposes S-Agent 2.0, a spatial reasoning agent that equips off-the-shelf general coding agents with a command-line perception toolkit, designing interfaces around acquiring scene evidence, writing to persistent scene memory, and interpreting readings in plain language, and reports evaluation results on MMSI-Bench and ReVSI. On MMSI-Bench and ReVSI, the toolkit improves the open-source model Qwen3.8-27B by 13.9 and 20.5 points respectively compared with the same agent without tools, and improves the closed-source model Gemini-3.8-Flash by 11.2 and 9.6 points respectively. The team states that as the base model becomes stronger the tool gains narrow, but on ReVSI the gain of the comparison method SpatialClaw drops from 13.2 to 2.7, while it maintains 9.6; S-Agent 2.0 can also run on different coding agent frameworks such as Codex and Claude Code.