ACT3D tops Gemini 3.6 Flash on six 3D reasoning benchmarks

ACT3D, a vision-language model fine-tuned with reinforcement learning, averages 53.2% accuracy across six 3D benchmarks and surpasses proprietary Gemini 3.6 Flash, according to a new arxiv preprint from Yijiang Li, Bingyang Wang, and Nuno Vasconcelos. It does this while seeing one image at a time.

The design is the point. Instead of feeding the model multi-view images, video, or an explicit 3D representation, the authors treat the scene as an environment the VLM interacts with through actions like 'move left' and observations rendered from the chosen viewpoint. 'Each rollout begins from an initial first-person view and a visual question,' the paper reports. 'The model can then iteratively reason, navigate, and reorient its viewpoint before committing to a final answer.'

RL does the teaching. After an initial trajectory-supervised finetune, the authors apply reinforcement learning with rewards 'driven primarily by final-answer correctness.' There is no direct supervision over which viewpoints to visit. 'The model must learn an effective exploration policy solely from the extent to which its interactions improve performance,' the abstract says.

The scaling pattern is what the authors call adaptive, question-conditioned exploration: a 0.6% gain over the base model on questions answerable from the initial view, climbing to 16.3% on questions requiring extensive exploration. ACT3D also beats existing open-source 3D methods while using 50-70% fewer input frames, and generalizes to unseen domains with a 5.9% average lift over its supervised initialization.

The abstract does not name the six benchmarks, does not give per-benchmark numbers, and does not report Qwen's parameter count or any comparison beyond Gemini 3.6 Flash.