SJTU's SpatialBrain Lifts Frozen GPT-5.6-sol to 72.6 Spatial Avg

Shanghai Jiao Tong and Zhejiang University's SpatialBrain, a framework that keeps the vision-language backbone fixed, raises the unweighted mean on five spatial benchmarks from 67.5 to 72.6 against direct answers from the same GPT-5.6-sol backbone, improving on four and tying on the fifth.

The largest gain lands on MindCube at 20.4 percentage points. On that benchmark, overall errors fall from 344 to 130, a 62.2% reduction. On MMSI the team reports 33 fewer errors overall, though the direction of change varies: "positional and attribute errors decrease, with no uniform reduction across all task types," the paper notes. Evaluated benchmarks are MMSI, MindCube, OmniSpatial, SpaceDG, and SAT, at 1,000, 1,050, 1,533, 1,000, and 150 items.

SpatialBrain stacks three modules around the frozen backbone. 3DFusionMemory fuses current-scene observations into queryable spatial relations; a Spatial Skill Library stores reusable tool procedures; Spatial Cognition associates task conditions with experience about when and how to use memory and skills. "The model weights remain fixed; the external experience available to the controller changes," the authors write. Completed trajectories are abstracted and consolidated between task blocks without reference answers.

The component ablation on MindCube's 1,050 items shows the full configuration at 87.6, versus 76.7 without the Skill Library, 75.2 without 3DFusionMemory, and 78.8 without Spatial Cognition, all above direct GPT-5.6-sol at 67.2. The authors caveat that removing 3DFusionMemory also disables memory-based answer checks, "so it does not isolate storage alone." They also flag that published baselines in their main table differ in protocols and serving precision.

Saved 3D objects occur in 50.6% of MMSI tasks and 37.5% of SpaceDG tasks, compared with 21.0%, 29.2%, and 18.0% on MindCube, OmniSpatial, and SAT, reflecting which inputs produce spatial state at all. The framework lands in a visible run of agentic-spatial work; USTC's SERA harness-editing framework for VLM planners crossed our tracker the same day.