[ICML 2025] "SepLLM: Accelerate Large Language Models by Compressing One Segment into One Separator" - View it on GitHub