SEEING ONCE IS ENOUGH? - Online Geometry-Aware Token Pruning for 3D Question Answering

Sep 21, 2026

Uploaded by Theodore

Recent Multi-modal Large Language Models (MLLMs) excel at 2D visual question answering but face major hurdles in 3D question answering due to multi-view requirements and high inference cost. Existing approaches depend on offline preprocessing of all frames or specialized compression strategies. This work introduces a first online, training-free, geometry-aware token pruning method that projects each posed RGB-D frame into a shared voxel space using depth and camera pose to prune redundant tokens. It reduces token usage by up to 50% while improving performance on ScanQA, SQA3D, and OpenEQA-HM3D with Qwen2.5-VL-7B and Qwen3-VL-8B.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages16

Reading Time40 min

AccessFree to download

Abstract

Introduction

Related Work

3D MLLMS

arXiv :2607 .04079v 1 [ cs .CV] 5 Jul 2026 SEEING ONCE IS ENOUGH? ONLINE GEOMETRYAWARE TOKEN PRUNING FOR 3D QUESTION ANSWERING Ruei-Chi Lai 1 Bolivar Solarte2 Chin-Hsuan Wu3 Yi-Hsuan Tsai4 Min Sun 1 1National Tsing Hua University 2Industrial Technology Research Institute ITRI 3University of Toronto 4Atmanity Inc. 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 tokenpruning 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. (a) (b) • non-overlap • overlap (c) Vision-Language Model kept tokens Pruned tokens Figure 1: Online token pruning. (a) Sequentially received posed RGB-D frames are voxelized into a shared 3D space. (b) Overlapped regions across frames are identified through voxel comparison.(c) Redundant visual tokens corresponding to these regions are pruned prior to being fed into the VLM, enabling substantial reduction of token usage without compromising performance. 1 INTRODUCTION Recent advances in Multi-modal Large Language Models (MLLMs) have demonstrated remarkable capabilities in 2D visual question answering (Bai et al., 2025; Team, 2025; Wang et al., 2025; Liet al., 2024; Achiam et al., 2023; Comanici et al., 2025) . However, extending these models to 3D question answering remains challenging. Unlike 2D tasks that rely on a single image, 3D understanding requires processing multiple frames to obtain sufficient spatial information, which greatly increases computational costs and token usage. Existing efforts have explored two main directions to address 3D question answering. First, several works inject 3D priors into MLLMs through specialized 3D modules to enhance spatial reasoning capabilities (Zheng et al., 2025; Xu et al., 2024b; Huang et al., 2025b; Chen et al., 2024a; Tanget al., 2024; Zhu et al., 2025; Hong et al., 2023; Zhang et al., 2024) . However, these approaches face significant limitations, requiring specialized 3D modules, extensive 3D-language paired datasets, and costly retraining pipelines, which restrict their practical application. Another line of research aims to reduce visual token usage through strategic frame selection or token compression (Huang et al., 2025a; Zheng et al., 2025; Wu et al., 2025; Xu et al., 2024a) . Although these approaches effectively reduce computational costs and maintain competitive performance, they require access to all images in the scene before inference, making them unsuitable for online tasks such as embodied AI, robotic navigation, or real-time scene understanding, which demand sequential, frame-by-frame processing. This raises the question: How can we lower computational cost while preserving online processing? To this end, we propose the first online, training-free, geometry-aware token pruning method (Figure 1) . Our approach leverages depth information and camera poses to project each input frame into a shared 3D voxel space, enabling the trackin

FAQs

Why is 3D question answering more costly than 2D?

3D understanding typically requires processing multiple frames to capture sufficient spatial information, which increases computational cost and visual token usage during inference.

What is the proposed online token pruning method?

The method projects each input frame into a shared voxel space using depth and camera pose, identifies spatially overlapped regions across frames, and prunes redundant visual tokens before feeding them into the vision-language model.

Does the method require additional training or offline preprocessing?

No. The approach is online and training-free, avoiding offline preprocessing that would require access to all images before inference.

Sep 21, 2026

Uploaded by Theodore