EvoCut

New methods drastically cut VLM visual tokens, boosting efficiency

Researchers have developed three new methods to significantly compress the visual tokens used by large vision-language models (VLMs), aiming to reduce computational overhead and improve inference speed. InfoMerge uses t…