The redundancy of visual tokens in multimodal large language models (MLLMs) motivates training-free token pruning techniques for efficient inference. Existing methods typically rank visual tokens by attention salience and retain only the most salient ones. However, salience-based pruning is susceptible to sink tokens and positional bias, overlooking the diversity of the selected set. Moreover, extracting attention maps incurs additional overhead. In this paper, we propose MI-Pruner, which derives Pointwise Mutual Information (PMI) based relevance scores in the projection space to reflect crossmodal relevance and intra-modal redundancy. Grounded in information theory, our approach identifies the most informative tokens by explicitly measuring feature-level dependency. Without reliance on internal attention maps or architectural modifications, MI-Pruner can be seamlessly applied to any pretrained MLLMs for inference acceleration. Extensive experiments on LLaVA1.5 and Qwen-series demonstrate that our approach achieves a favorable performance-efficiency trade-off. Specifically, we preserve ${>}97\%$ performance on LLaVA1.5-7B at a pruning rate of 88.9\%.
翻译:暂无翻译