Efficient Visual Token Compression for 3D Question Answering

3D Question Answering requires understanding a 3D scene and reasoning about natural language queries grounded in spatial context. Recent 3D vision-language models produce geometry-aware tokens by projecting 2D visual features into world coordinates, but generate thousands of tokens per scene, leading to substantial computational overhead. Existing token compression methods, mostly designed for 2D models, rely on semantic or attention-based selection and overlook the spatial structure of 3D tokens. In this paper, we propose a three-stage compression framework consisting of coarse voxelization, diversity-based anchor token selection, and spatially-constrained token merging. On three 3D-QA benchmarks, the proposed method retains 94.7% of the original performance with only 128 tokens and achieves 1.92× faster inference.