OTPrune：通过最优传输实现分布对齐的视觉令牌剪枝

OTPrune is a training-free framework that formulates visual token pruning as distribution alignment via optimal transport (OT). By minimizing the 2-Wasserstein distance between the full and pruned token distributions, OTPrune effectively reduces inference costs while preserving local diversity and global representativeness. Additionally, a tractable submodular objective is derived to enable efficient optimization, and theoretical proofs of its effectiveness are provided. This framework offers a novel approach for multi-modal large language models (MLLMs) in visual-language reasoning, particularly in addressing the issue of redundant visual tokens.