Paper review · October 22, 2026
S-Agent 2.0: Spatial reasoning with general-purpose coding agents
S-Agent 2.0: Spatial Reasoning with a General Coding Agent
By attaching a CLI perception toolkit with scene memory and spatial-language interpretation to off-the-shelf coding agents, the authors achieved higher performance on spatial reasoning benchmarks than dedicated agent systems.
Problem
Agentic VLMs that use 3D perception tools have greatly improved spatial reasoning performance. Yet the model often fails to produce the correct answer even when the tools return the right evidence. When the authors analyzed SpatialClaw Trajectories, most failures were not in the measurement stage. Instead, the model misread rotation signs or coordinate frames, or eyeballed the image to answer despite having measurements.
Method
General-purpose coding agents such as OpenHands were left as they are, and a command-line perception toolkit was added. recon merges multiple frames into a gravity-aligned, metric-scale 3D scene, and segment creates object masks using sentences written by the model as prompts. Results are stored in scene memory as named instances, and the model can also register virtual observers and the like with add-instance or remove incorrect instances with drop. Measurements such as distance, direction and motion are computed directly from memory and returned with a one-sentence interpretation that states the coordinate frame and reference object. Depth Anything 3, SAM 3, GeoCalib and others are used as fixed perception models.
Results
- With the Qwen3.8-27B backbone, the toolkit added +13.9 points on MMSI-Bench and +20.5 points on ReVSI over the same agent without tools.
- With Qwen3.8-27B, it scored 66.5 on MMSI-Bench and 82.3 on ReVSI, higher than SpatialClaw on the same backbone (62.6, 74.5).
- With Gemini-3.8-Flash, it scored 74.1 and 90.5, ahead of SpatialClaw (67.0, 79.7).
- Moving to a stronger backbone, SpatialClaw's ReVSI tool gain shrank from +13.2 to +2.7, while S-Agent 2.0 kept +9.6. It also worked in the Codex and Claude Code harnesses.
Limitations
The remaining failures the authors identify are in turning a question into the right measurement (model-to-tool grounding). Performance depends on the quality of the fixed external perception models. Evaluation is limited to question answering on public benchmarks, and the cost and latency from multiple tool calls were not addressed in the portion of the paper reviewed.
Why it matters
This shows that attaching well-designed CLI tools to general-purpose coding agents works better than designing a new dedicated agent loop, and that the gains can hold as backbones get stronger. The interface design in which tools return interpreted sentences instead of numbers is worth trying in agents for other domains.
Key terms
- Evidence loop
- The process of collecting evidence, remembering it and interpreting it through to an answer. If any of these links breaks, the answer is wrong even if the measurement is correct.
- Scene memory
- A persistent store holding a 3D point cloud and the position, extent and pose of each named object. The model only needs to remember names to retrieve measurements again.
- Coding agent harness
- An execution environment such as OpenHands, Codex or Claude Code that provides a shell, file system, code execution and a tool-calling loop.
Original
Authors: Rui Gong, Ruoyu Wang, Yalun Dai, Hao Li, Sidi Yang, Qihao Zhao, Hongyu Li, Zeyang Guan, et al. · Posted 2026-10-20 · License CC BY 4.0 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.