QbitAI · WeChatMay 7, 2026, 21:36 UTC+8

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: HKR-H/K/R all pass: 90% visual-token compression, no accuracy loss, and image-conditioned mapping give builders a testable cost-cutting mechanism. Zhejiang/Alibaba plus CVPR 2026 is strong research signal, not a platform-level product release.

Read the original ↗Export Markdown

Take

MetaCompress lands because it attacks the proxy, not the budget. Multi-turn VQA breaks prompt-conditioned pruning: the first question does not predict the third one. The paper’s sharpest number is the 1.71% overlap between optimally retained tokens and high-attention tokens. That is a nasty result for CLS- or prompt-attention pruning heuristics like FastV, and it gives the method a cleaner reason to exist than “we compressed tokens too.”

I still discount the “no accuracy drop” claim until the absolute benchmark tables, latency numbers, and training cost are checked outside the article. The body says MetaCompress beats FastV and PruMerge at 70% and 90% compression, with end-to-end cost near uniform downsampling. Code and arXiv are public, so this is testable. If it holds on LLaVA-NeXT-style multi-scale LVLMs, visual-token count stops being the dumbest bottleneck in multi-turn vision chat.

Source: QbitAI · WeChat · mp.weixin.qq.com
Published: May 7, 2026, 21:36 UTC+8 · May 7, 2026