Reproduce “LoopFill: Pipelined Multi-Token Speculative Decoding for Looped Transformers” 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

Siqi Zhu, Suozhi Huang, Yi Zhao, Jiaxuan You

LoopFill accelerates decoding in looped language models by combining retrieved suffix continuations with the model’s shallow-loop predictions in a shared draft tree. It advances candidate branches before full-depth verification, which alone determines emitted tokens. In greedy, concurrency-one tests across seven benchmarks, mean decode-latency speedups were 5.85× for Ouro-2.6B and 5.13× for Huginn-3.5B; these are not throughput or total-computation measures.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.