In vision-language models (VLMs), visual tokens usually consume a significant amount of computational overhead, despite their sparser information density compared to text tokens. To address this, ...
Some results have been hidden because they may be inaccessible to you
Show inaccessible results