Preprint
Aug 2026
RoRA: Role-Oriented Regional Allocation for Visual Token Pruning in MLLMs
RoRA is a training-free framework that casts visual token pruning as role-oriented regional evidence allocation, and consistently outperforms strong training-free baselines across LLaVA and Qwen-VL families, retaining most of the unpruned accuracy even at aggressive pruning ratios.
Qiyanhui Lu, Han Wu, Rongjia Xu et al.
· 0 citations