TokenGate: Adapting LLMs for pre-fill with learned skimming for ~11x speedup

Full paper: https://arxiv.org/html/2610.26754v1