Speculative decoding with any assistant model
Co-developed three methods, Dynamic Speculation Lookahead, Universal Assisted Generation and Token-Level Intersection, merged into Hugging Face Transformers; Token-Level Intersection is also in vLLM, where it is the only cross-tokenizer drafting path. Up to 2.8× faster generation with no change in output. First author on Universal Assisted Generation, which lets a small model draft for a large one even when their tokenizers differ.
ICML 2025 paperHF blog: UniversalHF blog: DynamicHF blog: TLI