S-Agent 2.0: Spatial Reasoning with a General Coding Agent

S-Agent 2.0. Spatial Reasoning with a General Coding Agent. By Rui Gong and eleven others, from Nanyang Technological University, the University of Hong Kong, and one other institution. Published on the twentieth of October. Its code repository is worked out: about two hundred and fifty source files, with a readme and a licence.

Spatial reasoning asks a model where objects sit in three-dimensional space, how they relate to each other, and how they move. Agentic VLMs have recently improved a great deal on these tasks, because instead of answering in a single pass they actively gather evidence with 3D perception tools. But the authors point to a stubborn failure. On many questions the tools return the right evidence, and the model still reaches the wrong answer. Between the model and its tools sits a gap, the place where raw readings should become usable conclusions, and nothing currently fills it.

S-Agent 2.0 is a spatial reasoning agent, and it is not built from scratch. It takes an off-the-shelf, general coding agent and equips it with a command-line perception toolkit. The effort goes into the interface between the two, which is designed around a single principle: help the VLM close the evidence loop. Getting a reading is only the start of that loop; the reading also has to be kept, and it has to be understood.

That loop has three steps. First, acquire evidence about the scene through the tools. Second, remember it, in a persistent scene memory, so that a reading taken earlier is still there when a later step needs it. Third, interpret each reading in plain spatial language, so the model receives a statement about where things are rather than a bare number it must decode. Built this way, the agent keeps the full flexibility of a coding agent, free to compose its own steps, while every reading it acts on stays grounded in the geometry the tools computed.

The tests use MMSI-Bench and ReVSI. Compared with the same agent running without tools, the toolkit lifts a strong open-source model, Qwen3.8-27B, by thirteen point nine points on MMSI-Bench and twenty point five points on ReVSI. A strong closed-source model, Gemini-3.8-Flash, gains eleven point two and nine point six points on those same two benchmarks.

On the open backbone, S-Agent 2.0 reaches sixty-six point five on MMSI-Bench and eighty-two point three on ReVSI. The specialised system SpatialClaw, on the same backbone, scores sixty-two point six and seventy-four point five.

With Gemini-3.8-Flash as the backbone, S-Agent 2.0 reaches seventy-four point one on MMSI-Bench and ninety point five on ReVSI. SpatialClaw, on that backbone, gets sixty-seven and seventy-nine point seven. So the lead holds on both benchmarks with both backbones.

Both systems gain less from tools as the backbone becomes more capable. But on ReVSI, SpatialClaw's gain shrinks far more, falling from thirteen point two points to two point seven, while the gain for S-Agent 2.0 holds at nine point six. The agent also works across different general coding agent harnesses, including Codex and Claude Code.

What sets this apart from simply adding more tools is where the effort goes. The perception was already returning the right numbers; S-Agent 2.0 spends its design on the step after, turning a reading into something the model can remember and reason with.