Brief

Researchers have developed a novel method for processing gigapixel whole slide images in vision language models by treating token reduction as a trainable sparsification problem. This approach, detailed in a new arXiv paper, allows the model to learn an optimal selection strategy for visual tokens, unlike previous methods that used non-trained downsampling or heuristic pruning. The proposed decoupled routing architecture and SparseLearn component enable gradient propagation through the pruning process, ultimately reducing the visual sequence to a sparse set of 32 tokens with minimal computational overhead during inference. This technique achieves high accuracy on benchmarks like SlideBench, offering an efficient paradigm for end-to-end gigapixel image reasoning. AI