Reproduce “Looped Reasoning Finishes Earlier Than You Think” 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
Sumin Park, T. Konstantin Rusch
QRR treats token-wise phase rotations as representational symmetries, so its task-relevant quotient state can converge while raw state continues moving. Its guarantee covers only the designed symmetry directions; other quotient dynamics may still oscillate. On Sudoku-Extreme, certified-root median convergence takes 122 layer-forwards, versus 234 for FPRM, with ground-truth-decoding roots found for 89.6% versus 90.4%.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.