Reproduce “Multi-Future Drafting: Mixture Block Drafters for Tree-Verified Speculative Decoding” 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
Haoyu Huang, Jialiang Kang, Linlin Yang, Xiao Yang, Baochang Zhang
MFD drafts multiple coherent continuations in parallel and merges their prefixes into one tree for verification, avoiding mixed paths from position-wise rankings. On Qwen3-4B at temperature zero, it achieved a five-task-mean decoding speedup of 7.32×, versus 6.66× for three-epoch D-PACE trained on the same 1.42-million prompts. This result does not establish fewer tokens or an advantage at high batch sizes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.