Reproduce “DLoop: Looped 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
Geonmo Gu, Byeongho Heo, HeeJae Jun, Yoohoon Kang, Sangmin Lee, Sangdoo Yun, Dongyoon Han
DLoop reduces target-model verification passes by letting a draft model propose token batches before checking them. A confidence gate decides when to continue, and loop-aware training teaches the drafter to use its hidden states across stages. With DSpark and Qwen3-4B, eight-benchmark average speedup rises from 3.77× to 4.76×; this trades draft passes for fewer target passes, not lower computation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.