Each Complexity Deserves a Pruning Policy
Hanshi Wang 1,2,3,5 , Yuhao Xu 3 , Zekun Xu 1,2 , Jin Gao 1 , 2 , 5 † {}^{1,2,5\text{\textdagger}} , Yufan Liu 1,2,5 , Weiming Hu 1,2,5,6 , Ke Wang 7 , Zhipeng Zhang 3,4 1 State Key Laboratory of Multimodal Artificial Intelligence Systems (MAIS), CASIA 2 School of Artificial Intelligence, Universi · 2025
Read Each Complexity Deserves a Pruning Policy by Hanshi Wang 1,2,3,5 , Yuhao Xu 3 , Zekun Xu 1,2 , Jin Gao 1 , 2 , 5 † {}^{1,2,5\text{\textdagger}} , Yufan Liu 1,2,5 , Weiming Hu 1,2,5,6 , Ke Wang 7 , Zhipeng Zhang 3,4 1 State Key Laboratory of Multimodal Artificial Intelligence Systems (MAIS), CASIA 2 School of Artificial Intelligence, Universi (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-23931.txt.
Description
Vision-language models read hundreds of image tokens when a handful might carry the answer. Hanshi Wang and 7 co-authors prune those tokens adaptively with AutoPrune, a training-free framework. Human vision explores broadly first, then narrows; their experiments show models do something similar. A fixed pruning schedule cannot fit inputs of differing complexity, so AutoPrune measures mutual information between visual and text tokens.
That signal is projected onto a budget-bound logistic retention curve shaped per task. Applied to LLaVA-1.5-7B, the method cut most visual tokens and three quarters of the compute while keeping nearly all accuracy. It also transferred to driving-oriented vision-language-action models.
Catalog record
| CREATOR | Hanshi Wang 1,2,3,5 , Yuhao Xu 3 , Zekun Xu 1,2 , Jin Gao 1 , 2 , 5 † {}^{1,2,5\text{\textdagger}} , Yufan Liu 1,2,5 , Weiming Hu 1,2,5,6 , Ke Wang 7 , Zhipeng Zhang 3,4 1 State Key Laboratory of Multimodal Artificial Intelligence Systems (MAIS), CASIA 2 School of Artificial Intelligence, Universi |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 40 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-23931