Reproduce “Multi-Draft On-Policy Distillation for Diffusion Drafters” 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
Petr Molodyk, Yuchen Zhu, Jaemoo Choi, Molei Tao, Yongxin Chen
MDraft trains a parallel drafter by combining verifier distributions from multiple sampled drafts of the same context, while leaving the drafter architecture and inference procedure unchanged. It addresses conflicting prefix-specific targets with blended targets and reverse KL training. On Qwen3-4B, its six-benchmark mean acceptance length is 5.903 at T=1, versus 5.726 for RKL; these results measure acceptance, not wall-clock speedup.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.