Found first: a primary source the press has not covered yet.
A team from UC San Diego and Georgia Institute of Technology has released ACT3D, an open vision-language model that scores 53.2% on six 3D spatial reasoning benchmarks, one percentage point above Gemini 3.6 Flash's 52.2%, while using 50-70% fewer input frames than existing open-source 3D methods. The work is described in a preprint submitted to arXiv on October 15, 2026.
What the source says
Yijiang Li, Bingyang Wang, and Nuno Vasconcelos built ACT3D on a Qwen3.5-9B base, first fine-tuning on supervised navigation trajectories and then applying RL with rewards driven solely by final-answer correctness, with no direct supervision over which viewpoints the model selects. The model receives one image at a time and decides where to look next as it moves through a scene. Evaluated on SPBench-SI, SPBench-MV, ARKitScenes, ScanQA, SQA3D, and MP3D, ACT3D gains 22.0 percentage points over its Qwen base. It also generalises to unseen domains, showing a 5.9 percentage point average improvement over a supervised baseline. The performance gap scales with exploration demand: on questions answerable from the initial view, the gain is 0.6 pp; on questions requiring extensive exploration, it is 16.3 pp.
Why it matters
Most 3D VQA approaches supply a model with many views at once, pre-solving the viewpoint problem through engineering rather than learning. ACT3D acquires adaptive viewpoint selection through RL, with the model learning when to explore and when to stop. The 50-70% frame reduction is not just an efficiency figure; it indicates the model is skipping redundant observations rather than accumulating them. An open 9B-parameter model matching and exceeding Gemini 3.6 Flash on this task class is a concrete data point on where open and closed systems now stand in embodied spatial reasoning.