LinMU achieves linear complexity for multimodal understanding models

Researchers have developed LinMU, a novel Vision-Language Model (VLM) architecture that achieves linear complexity, overcoming the quadratic complexity limitations of current models. This new design utilizes an M-MATE block, combining a state-space model with window attention, to process high-resolution images and long videos efficiently. Through a three-stage distillation process, LinMU matches the performance of existing models while significantly reducing processing time and increasing throughput, making advanced multimodal reasoning more accessible. AI

影响 Enables more efficient processing of high-resolution images and long videos, potentially leading to wider adoption of advanced multimodal reasoning.

排序理由 This is a research paper detailing a new model architecture and training methodology. [lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv cs.CV 阅读 →

• Hongjie Wang
• LinMU
• Local-Swin
• LongVideoBench
• LoRA
• TextVQA
• Video-MME

• 论文
• 模型发布

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →