Reproduce “Not All Loops Cost the Same: Phase-Dependent Depth for Serving 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

Zhixu Du, Yuhan Cheng, Hai "Helen" Li, Yiran Chen

LoopDuet reduces recurrent prompt computation in looped transformers, retaining full loop depth for first-token prediction and generated tokens. It removes finished prompt positions from batched work and adjusts prefilling depth as queues grow. On Huginn HotpotQA, (14,8) achieved 30.33 normalized exact match and 40.35 F1, versus 24.29 and 33.22 at 97.3% of uniform depth-eight throughput; results are setting-specific.

Summary by alphaXiv.

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