Reproduce “Classical Energies Learn to Reason 1,000x Faster than Looped Models” 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
Benjamin Hoover, Hugh Blayney, James C. R. Whittington
CRAM (Compact Reasoning Associative Memory) learns a pairwise energy over candidate puzzle boards and searches for low-energy solutions rather than repeatedly predicting digits. It guides cell-by-cell search while keeping clues fixed. It solves 99.996% of 412,786 Sudoku-Extreme test puzzles, with 282-fold fewer inference FLOPs than EBR at matched accuracy; these structured-puzzle results do not establish general reasoning ability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.