Reproduce “Energy-Based Reasoners: Your Recurrent Reasoner Should Be an Energy-Based Model” 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
William Xingxu Chen, Alexi Gladstone, Zeyu Michael Li, Heng Ji, Yilun Du
EBR gives recurrent reasoning a learned energy score to guide hidden-state updates toward lower energy; its gradient also ranks candidate answers and signals when to stop. On Sudoku-Extreme, it reaches 99.92% exact accuracy at 0.61 TFLOPs per puzzle, versus EqR's 99.8% at about 210 TFLOPs, comparing inference compute, not elapsed time. Evidence covers structured algorithmic benchmarks, not general language reasoning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.