Reproduce “Decoupled Marginal Sharpening for Training-Free Inference-Time Scaling” 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

Weixin Wang, Wei Deng, Anderson Schneider, Yuriy Nevmyvaka, Pan Xu, Andrew Bennett

DMS separates sharpening of reasoning paths from sharpening of the answer support pooled across those paths, without retraining or an external verifier. It samples weighted reasoning paths, aggregates model support for candidate answers, then favors answers with greater pooled support. Across 12 model–benchmark settings, it reports highest mean pass@1 among compared methods; on AIME24/25, answer sharpening alone matches its pass@1.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.