S-Agent 2.0 uses a general coding agent for spatial reasoning, topping five benchmarks
ropedia-ai · hf · 2026-10-21
Researchers from NTU, HKU and Ropedia present S-Agent 2.0, which integrates persistent scene memory and plain-language interpretation into a general coding agent with 3D perception tools for spatial reasoning. Its reasoning loop plans evidence (target, viewpoint, relation), builds the scene, binds references, and queries evidence. It beats baselines including SpatialClaw, SpaceTools and Qwen3.5-275B across MMSI, MindCube, SPAR, ReVSI and MMSI-Vid (66.5 on MMSI, +14.1pp over Qwen3.5-275B). Code is open-sourced.
More from coding & agent
- Why agents need an 'Agent GitHub' for versioning, permissions, and memory — yangyi · 2026-10-21
- Kapitsa 0.2's KFlash tree speculative decoding beats Splash 1.4.0 with 63 tok/s on Qwen 27B — DotaMate · 2026-10-21
- Dev predicts 2027 will be the most fun year as vibe-coded games take off — TAbrodi · 2026-10-21
- Hamel Husain: public benchmark quality is abysmal — broken graders, sloppy harnesses — HamelHusain · 2026-10-21
- Dev jokes Claude is never blocked from making mistakes, only from fixing them — cjimti · 2026-10-21
- Compile Your LLM Workflow Into Small Model Weights: A LoRA Hypothesis Worth Testing — arpit_bhayani · 2026-10-21