A new pruning technique lets AI models answer questions about 3D scenes in real time while cutting token usage by up to 50%.
Researchers published a method that trims redundant image tokens before they reach the language model in multi-modal 3D question-answering pipelines. Rather than preprocessing every camera view offline, the system maps each incoming frame into a shared 3D voxel grid using depth data and camera position, then discards tokens from regions already covered by prior frames. The approach requires no retraining and plugs into existing models without modification. Tests on Qwen2.5-VL-7B and Qwen3-VL-8B showed benchmark improvements on ScanQA, SQA3D, and OpenEQA-HM3D.
Most efforts to make 3D scene understanding cheaper have required selecting frames in advance or merging tokens offline - both assume the full scene exists before inference begins. An online, frame-by-frame approach opens the door to real-time 3D question answering in robotics or augmented reality, where the world arrives incrementally. Cutting token count also reduces inference cost directly, which is where running these models gets expensive at scale.
The key number is an up-to-50% reduction - a ceiling, not a guarantee, and one worth remembering before those benchmark figures travel.