Researchers propose Entropy-Aware Dense Pruning (EADP), a framework designed to address the failure of existing visual token pruning methods to preserve critical cues under dense instructions. The approach reformulates pruning as a structured compression problem by leveraging statistical entropy to filter textual noise and casting token selection as a submodular maximization problem with a spatial prior.

  • EADP quantifies and filters out textual noise using statistical entropy to yield robust instruction relevance scores.
  • Token selection is cast as a submodular maximization problem with a spatial prior to ensure holistic and non-redundant visual representation.
  • The method improves the accuracy-efficiency trade-off of Vision-Language Models (VLMs) under strict token budgets.
  • EADP achieves state-of-the-art performance on challenging multimodal benchmarks while preserving fine-grained visual cues.

The authors consider this important because it robustly preserves fine-grained visual cues that are often lost by naive Top-K selection, thereby improving the overall efficiency and accuracy of VLMs.