Reproduce “Sequent-Prover: Training Agents for Formal, Checkable SMT-based Reasoning” 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
Debargha Ganguly, Weicong Chen, Mohsen Hariri, Vikash Singh, Sreehari Sankar, Shouren Wang, Wang Yang, Biyao Zhang and 8 more
The authors train language models to translate natural-language questions into SMT-LIB programs and revise them using Z3’s feedback. They filter solver-supported trajectories for SFT and DPO. On a 37-benchmark evaluation, 4B DPO raises average accuracy from 0.856 to 0.970 and pass^4, the share with four correct samples, from 0.641 to 0.917. But solver checks formalizations, not whether they preserve question meaning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.