Shun-Jie Wen

Centering before Pruning: Lightweight Geometry Correction for Diversity-Based Visual Token Pruning in LVLMs

The proposed Cen-Prune is a lightweight, plug-and-play correction, which measures subset diversity using centered cosine similarity while retaining raw-space distinctiveness as a complementary token-wise preference and provides robust improvements in overall performance across existing diversity-based pruners.