Researchers have developed LiLiCorr, a model that improves the performance of speculative decoding in language models. This technique accelerates inference by drafting future tokens and verifying them in parallel. However, previous methods have produced individually plausible but jointly incoherent results. LiLiCorr addresses this issue by correlating the marginal distributions of per-position tokens, allowing it to capture the joint structure of the block without materializing the full joint distribution. The model is co-trained with a drafter to learn proposing candidates that correlate into longer accepted sequences, resulting in improved acceptance lengths and throughput. --- Why it matters: This matters because speculative decoding can significantly accelerate language-model inference, but previous methods have limitations. LiLiCorr's ability to capture block-level structure and improve coherence can lead to better performance in real-world applications, making it an important advancement for researchers working on efficient language models. Source: https://arxiv.org/abs/2608.20530