Technique to decrease the number of tokens processed by a model, typically by compressing or filtering visual information.