Reproduce “DForkTree: Accelerating Speculative Decoding with Fork-Aware Block Diffusion Draft Trees” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhicheng Zheng, Cheng Zhang, Chenkai Tang
The authors introduce DForkTree, a tree-based speculative decoding method that adjusts a block drafter’s candidate scores using target-generated n-gram statistics from each branch’s preceding tokens, without another drafter pass. In tests, it reaches 5.47× speedup over autoregressive decoding on Qwen3.5-4B and MATH-500 at temperature 0, within batch-size-one evaluations of block-diffusion drafters.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.