New method prunes tokens for efficient 3D question answering

Researchers have developed a novel online token-pruning method designed to enhance the efficiency of multi-modal large language models (MLLMs) in 3D question answering tasks. This approach projects input frames into a shared voxel space, identifying and pruning spatially overlapped regions to reduce redundant image tokens before they enter the language model. The method, which requires no additional training, can decrease token usage by up to 50% and has shown improved performance on benchmarks like ScanQA, SQA3D, and OpenEQA-HM3D when applied to models such as Qwen2.5-VL-7B and Qwen3-VL-8B. AI

IMPACT This method could significantly reduce computational costs for 3D AI applications, enabling more efficient processing and potentially wider adoption.

RANK_REASON The cluster contains a research paper detailing a new method for 3D question answering. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• OpenEQA-HM3D
• Qwen2.5-VL-7B
• Qwen3-VL-8B
• ScanQA
• SQA3D

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →