UBTree Parallel Tree Drafting via Unigram and Bigram Models
October 1, 2026
UBTree improves speculative decoding performance in high-entropy distributions by coupling a unigram proposer with a bigram selector. This method constructs drafting trees using a renormalized KL objective on high-temperature data to increase draft diversity without increasing latency.
HOW THIS AFFECTS YOU
●
builderThis could reduce inference latency in production models by improving speculative decoding efficiency.
●
researcherYou can leverage tree-native training to broaden supervision beyond the greedy path.