Reproduce “Tree-Ising: Repairing Language Models with Bit-Level Ising Edits” 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
Hadi Salloum, Ammar Ali
Tree-Ising searches for sparse exact changes to model weights to reduce reasoning loops while preserving productive outputs. It contrasts looping and successful traces, then solves row-wise binary optimization problems over admissible weight edits. On Qwen3.5-0.8B, it scores 54.6% on MATH-500@4, versus 51.7% for an agent edit in a shared evaluation; the comparison is not dose-matched.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.