A new method for pruning visual tokens in vision-language models has been proposed. The E2S-Pruner framework estimates the reliability of each attention head and represents visual tokens as important, unimportant, or uncertain. It then uses Dempster-Shafer evidence theory to fuse evidence from multiple network layers and promote coverage of distinct image regions. The authors claim that their method retains 90-98% of the original model's performance while improving throughput by up to 2.09x. --- Why it matters: This matters because pruning visual tokens can significantly reduce the inference latency and GPU memory overhead of vision-language models, making them more practical for real-world applications. Source: https://arxiv.org/abs/2608.23253