Reproduce “Scaling to Tens of Thousands of Test-Time Iterations with Loop-Native Attention Residuals” 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
Pengxiang Li, Dilxat Muhtar, Di He, Guinan Su, Lu Yin, Shiwei Liu
InfiLoop adds test-time depth to recursive Transformers by reusing a shared block while retaining a learned, decaying summary of earlier candidate states and weighting new ones by content. In Sudoku-Extreme, the 7M-parameter model reaches 90.9% exact accuracy at 1,872 executed layers, compared with FPRM's 90.1% at 3,744. These are layer-evaluation counts, not runtime measurements, and results cover symbolic tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.