Seeing Once is Enough? Online Geometry-Aware Token Pruning for 3D Question Answering

2026-07-04

AI Summary

This work addresses the high computational overhead and offline preprocessing requirements of existing 3D visual question answering methods, which hinder online deployment. The authors propose an online, geometry-aware image token pruning approach that enables efficient 3D reasoning without any training or offline preprocessing. By projecting multi-view frames into a shared voxel space using depth and camera poses, the method geometrically aligns views and dynamically prunes redundant tokens based on spatial overlap detection. This technique seamlessly integrates with off-the-shelf multimodal large language models such as Qwen2.5-VL-7B and Qwen3-VL-8B, achieving comparable or improved performance on ScanQA, SQA3D, and OpenEQA-HM3D benchmarks while reducing token usage by up to 50%, thereby significantly lowering memory consumption and inference latency.

Abstract

Recent Multi-modal Large Language Models (MLLMs) have demonstrated remarkable performance on 2D question answering tasks. However, extending these models to the 3D question answering remains challenging, as they typically require multiple views of the scene, which incurs substantial computational cost at inference. To mitigate this issue, existing solutions rely on strategic frame selection or token-merging algorithms that require preprocessing in advance all frames of the scene, i.e., an offline fashion. In contrast, we propose the first online token-pruning method that can be integrated seamlessly with current MLLM models for 3D question answering tasks, without additional training and with lower memory usage.Our key insight is to project each input frame into a shared voxel space using depth information and camera pose, identifying spatially-overlapped regions across frames and selectively pruning redundant image tokens before they enter the language model. Our method enables efficient online processing while reducing up to 50% of token usage. We apply this approach to Qwen2.5-VL-7B and Qwen3-VL-8B, demonstrating improved performance on the ScanQA, SQA3D, and OpenEQA-HM3D benchmarks.

Problem

• 3D question answering
• token pruning
• online processing
• multi-modal large language models
• computational efficiency

Innovation

• online token pruning
• 3D question answering
• geometry-aware
• voxel space projection
• multi-modal LLM