Reproduce “Numina Fuse: Interactive Autoformalization in Lean 4” 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
Justin Asher, Yichuan Cao, Ruichen Qiu, Pietro Pellegrino, Jia Li
Numina Fuse links informal blueprints to Lean declarations in a shared workspace, where users can inspect and redirect agent work. Lean checks proofs, not whether statements match intended mathematics. In a 52-lemma study, Doubao Seed 2.1 Turbo solved 52/52 at $0.080 per solved lemma, while Claude Opus 4.8 solved 49/52 at $0.526. This does not measure Fuse’s overall productivity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.