According to Stanford AI Lab via X, Token-Gated Language Models learn to skim long-context inputs and conditionally forward salient tokens, delivering up to 11x faster prefill compared with dense processing.