AI summaryⓘ
The authors studied how to reduce the cost of large vision-language models by selecting fewer important visual tokens. They found that simply using the usual similarity measures doesn't distinguish tokens well because most similarities are positive and close to each other. By centering the token features before measuring similarity, they revealed more detailed differences, but this alone made performance worse. They discovered that raw features help pick globally distinctive and meaningful tokens, so they designed a new method called Cen-Prune that combines centered similarity for diversity with raw features for distinctiveness, improving token selection without extra cost. Their experiments show this approach boosts performance across various visual tasks.
large vision-language modelsvisual tokenscosine similaritytoken pruningfeature centeringsubset diversitytoken distinctivenessinference costvision-language benchmarksCen-Prune
Authors
Shunjie Wen, Jaeyeon Lee, Dong-Wan Choi
Abstract
Large vision-language models (LVLMs) incur substantial inference costs due to their long and highly redundant visual-token sequences. Diversity-based pruning mitigates this cost by selecting token subsets based on pairwise cosine similarity. We find, however, that similarities between raw visual tokens are strongly concentrated in the positive range, limiting their ability to distinguish non-redundant tokens. A natural way to improve this resolution is to center token features before computing cosine similarity. Centering indeed reveals a substantially richer pairwise structure, yet unexpectedly degrades pruning performance when used alone. We show that this apparent contradiction arises because the raw geometry does more than represent pairwise diversity: it also implicitly favors globally distinctive tokens, which tend to contain semantically informative content. Centering better resolves subset diversity but loses this useful token-wise preference, revealing that diversity and distinctiveness are entangled in the raw geometry. Based on this analysis, we propose the \textbf{Cen}tered Geometry \textbf{Prune}r (Cen-Prune), which measures subset diversity using centered cosine similarity while retaining raw-space distinctiveness as a complementary token-wise preference. This lightweight, plug-and-play correction leaves the underlying selection mechanism unchanged and incurs negligible computational overhead. Extensive experiments across multiple image- and video-understanding benchmarks and LVLM architectures demonstrate that Cen-Prune provides robust improvements in overall performance across existing diversity-based pruners.