Reproduce “Vibe-Coded and Tuned: A State-of-the-Art SMT Solver for QF-LRA” 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
Mikoláš Janota, Jan Jakubův
Primo is a linear-arithmetic constraint solver built by an LLM coding agent under human supervision. The authors used literature-guided implementation, benchmarking, profiling, fuzzing, and tuning. At a 30-second cutoff, tuned primo⋆ solved 1,599 of 1,753 QF-LRA instances, versus 1,578 for default OpenSMT. The advantage is limited to reported comparisons against default configurations, not tuned competitors.
Key insights
- The strongest gains came from configuration choices, particularly model-based phase selection and monotone variable elimination, rather than from introducing a new linear-arithmetic decision procedure.
- Primo⋆ showed especially strong behavior on the LassoRanker family and on satisfiable or unknown-status instances, while OpenSMT retained an advantage on some unsatisfiable cases.
- LLM-assisted implementation can support substantial solver engineering when paired with human direction, systematic testing, profiling, documentation, and automated tuning; it does not by itself establish correctness or broad generalization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.