arxivcs.CVcs.AIcs.LG2026-07-05
Seeing Once is Enough? Online Geometry-Aware Token Pruning for 3D Question Answering
Ruei-Chi Lai, Bolivar Solarte, Chin-Hsuan Wu, Yi-Hsuan Tsai, Min Sun
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 substant…