SDSAT: Accelerating LLM Inference through Speculative Decoding with Semantic Adaptive Tokens Paper • 2403.18647 • Published Mar 27, 2024