Reproduce “Rotate the Loop: Blockwise Conformality 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
Alexey Zemtsov, Albina Klepach, Vadim Lomshakov, Andrei Polubarov, Nikita Lyubaykin, Alexander Derevyagin, Vladislav Kurenkov, Alexander Nikulin
The authors modify looped transformers’ transition map so a shared block can read information across repeated passes. The map rotates and contracts coordinate-aligned groups at separate rates, preserving comparisons within each group. On A₅ state tracking at T=32, it reaches 0.894 ± 0.052 last-token accuracy, versus 0.692 ± 0.097 for the learned-diagonal baseline; G=2 is not best at every group count.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.