LearnPruner: Rethinking Attention-based Token Pruning in Vision Language Models

Name: LearnPruner: Rethinking Attention-based Token Pruning in Vision Language Models
Uploaded: 2026-10-11T06:22:13.328626+00:00
Duration: 796 s
Description: LearnPruner is a two-stage framework to prune Vision-Language Models by removing redundant visual tokens and then filtering text-irrelevant tokens in the LLM's middle layer. It uses learnable modules to predict token importance, achieving 95% of original performance while using only 5.5% of vision tokens and providing significant inference speedup.

LearnPruner is a two-stage framework to prune Vision-Language Models by removing redundant visual tokens and then filtering text-irrelevant tokens in the LLM's middle layer. It uses learnable modules to predict token importance, achieving 95% of original performance while using only 5.5% of vision tokens and providing significant inference speedup.

Open the full episode