Combating Textual Noise and Redundancy: Entropy-Aware Dense Visual Token Pruning

Vision-language models (VLMs) process images as many small "visual tokens," and pruning redundant ones speeds up inference—but existing pruning methods often discard important details when instructions are dense or fine-grained. This paper identifies two causes: noisy textual signals corrupting relevance scoring, and fragmented feature selection. Their proposed method, EADP, uses statistical entropy to filter noise and reframes token selection as a submodular optimization problem ensuring diverse, non-redundant coverage. This improves the accuracy-efficiency tradeoff for VLMs, which is valuable for deploying multimodal AI systems—like visual assistants or document analyzers—under strict computational budgets. Authors: Xuehui Wang, Xuankun Yang, Wei Shen Paper: https://arxiv.org/abs/2607.02484v1