What it is
A VLM learns to answer 3D questions by actively moving through a scene one view at a time, then answering after RL-trained exploration and reasoning. Instead of feeding full 3D, video, or multiview inputs, it keeps the 3D scene external and trains a single-image VLM to explore via actions. RL uses mainly final answer correctness, with no trajectory labels for where to look.
Key results
- 53.2% avg accuracy on 6 benchmarks, +22.0% over base Qwen
- Beats Gemini 3.6 Flash by 1.0% average accuracy
- Uses 50-70% fewer input frames than prior open-source 3D methods
Outcome-based RL can teach VLMs not just to reason from observations, but to acquire the right observations for 3D tasks.