Domino: Decoupling Causal Modeling from Autoregressive Drafting in Speculative Decoding

Researchers introduced Domino, a speculative decoding framework for faster LLM inference.

Domino splits drafting into a parallel backbone and a lightweight prefix-aware correction head, backed by a stabilizing training curriculum.

On Qwen3 models, it reported up to 5.49x end-to-end speedup with Transformers and 5.8x throughput gain with SGLang.

The approach aims to preserve causal dependency modeling while keeping block-parallel drafting efficient, improving latency and throughput.