V-CoLA

Headlines · 1

1. Oct 7
   V-CoLA: training-free vision token compression for linear-attention VLMs

   V-CoLA is a training-free vision token compression framework designed for hybrid architectures that use linear attention (the paper cites Qwen3.5 as an example). It introduces a uniqueness-aware importance criterion to pick critical vision tokens and an adaptive token merging strategy to compress them, with all components tuned to stay compatible with the chunk-wise parallelism of linear attention. The authors report reaching 99.5% of original performance with only 50% of vision tokens, over 88% with as few as 12.5%, and 1.86x to 6.15x prefill speedups across multiple benchmarks. They also note that prior softmax-attention-based attention or similarity compression methods degrade noticeably in this regime.

   Hugging Face · Papers · 🔥 0

Experience and discussion from the community

Share my V-CoLA experienceAsk about V-CoLA