S-Agent 2.0 turns coding agents into spatial reasoners, hitting 66.5 on MMSI-Bench
liuziwei7 · x · 2026-10-21
Researchers from NTU and collaborators released S-Agent 2.0, a spatial reasoning agent that equips an off-the-shelf general coding agent with a command-line perception toolkit. The design centers on closing the "evidence loop": acquire scene evidence, store it in persistent scene memory, and interpret each reading in plain spatial language so conclusions stay grounded in tool geometry.
Key results:
- The toolkit lifts open-source Qwen3.8-27B by +13.9 and +20.5 points on MMSI-Bench and ReVSI versus the same agent without tools; closed-source Gemini-3.8-Flash gains +11.2 and +9.6
- With the open backbone, S-Agent 2.0 reaches 66.5 on MMSI-Bench and 82.3 on ReVSI, beating the specialized SpatialClaw on the same backbone (62.6 / 74.5)
- Paper: arXiv 2610.28354
Related event: S-Agent 2.0 Turns General Coding Agents into Spatial Reasoners(2 posts)→
More from coding & agent
- Reverse-engineering Futurama game took a month and $1,266 of Claude Code Opus — JasonBotterill · 2026-10-21
- 10 open-source repos for building AI research agents, from GPT Researcher to Docling — Shruti_0810 · 2026-10-21
- Claude vs Codex for browser automation: Codex acts, Claude balks at trivial steps — Capaj · 2026-10-21
- Anchor SDK: an open-source TypeScript agent harness with sandboxes, memory, and traces — Elegant_Fish_3822 · 2026-10-21
- What actually broke moving agents from demo to production: 5 lessons — No-Age-3362 · 2026-10-21
- Vanguard cut legacy code research from 7 days to 15 minutes by exposing mainframe via MCP — AI Engineer · 2026-10-21