Researchers have introduced DominoTree, a new method for speculative decoding in large language models. It builds on the Domino drafter by adding a GRU-based causal correction to each draft token's distribution. This makes the draft tree path-dependent, unlike previous methods like DDTree. The team evaluated DominoTree on eight benchmarks and found it outperformed other decoders, including Domino, in terms of accepted length and throughput. --- Why it matters: This matters because speculative decoding is a crucial technique for accelerating large language model inference, and improving its efficiency can lead to faster and more accurate results. Engineers working with LLMs will be interested in the performance gains achieved by DominoTree compared to other decoders. Source: https://arxiv.org/abs/2607.08642