Skip to content
QbitAI · WeChat

Zhejiang University and Alibaba MetaCompress reaches 90% token compression for multi-turn VQA

只看图片就能学会压缩Token!浙大&阿里新框架多轮VQA压缩率90%,精度不掉|CVPR 2026

Zhejiang University and Alibaba proposed MetaCompress, a learned token-compression framework that generates a compression mapping from the input image alone for multi-turn VQA. The article says it can remove 90% of visual tokens while preserving accuracy, and reports only 1.71% overlap between optimally retained tokens and high-attention tokens.

Why it matters: Zhejiang University and Alibaba let a model decide which visual tokens to keep from the input image alone, cutting 90% of tokens in multi-turn VQA with no accuracy loss.

Read the original ↗Export Markdown