TIDE. Dynamic per-token early exit for LLM inference. Skip layers tokens don't need