Research map / Large language models / Agents, code and math
From LLM-Generated Conjectures to Lean Formalizations: Automated Polynomial Inequality Proving via Sum-of-Squares Certificates
Ruobing Zuo, Hanrui Zhao, Gaolei He, Zhengfeng Yang, Jianlin Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add large-scale synthetic SOS cold-start SFT before RL (ablation) | NSPI SOS conjecturer · PolyIneq-Synth | pass rate: 14.76% -> 25.95% | Improved |
| add curriculum-based GRPO after SFT (ablation) | NSPI SOS conjecturer · PolyIneq-Synth | pass rate: 21.19% -> 25.95% | Improved |
| add curriculum partitioning to GRPO (ablation) | NSPI SOS conjecturer · PolyIneq-Real | pass rate: 42.16% -> 43.14% | Improved |
| add curriculum partitioning to GRPO (ablation) | NSPI SOS conjecturer · PolyIneq-Synth | pass rate: 24.29% -> 25.95% | Improved |
| SFT on synthetic data versus base model | SOS conjecturer · PolyIneqBench overall | success rate: 15.9% -> 25.3% | Improved |
| add progressive RL stages after SFT | SOS conjecturer · PolyIneqBench overall | success rate: 25.3% -> 29.3% | Improved |
| use NSPI instead of LIPS | hybrid prover · PolyIneq-Real | pass rate: 43.14% vs 67.65% | Hurt |
| use NSPI instead of LIPS | hybrid prover · PolyIneq-Synth | pass rate: 25.95% vs LIPS (2.87-fold improvement) | Improved |
| use NSPI instead of symbolic Maple/Z3 | low-variable problems · PolyIneqBench n=3 | pass rate: 44.1% vs 97.6% (Maple) | Hurt |
Closest papers by meaning
- Proving Olympiad Inequalities by Synergizing LLMs and Symbolic Reasoning (ICLR 2025 · similarity 0.90)
- Solving Inequality Proofs with Large Language Models (NeurIPS 2025 · similarity 0.88)
- Automated Proof of Polynomial Inequalities via Reinforcement Learning (CVPR 2025 · similarity 0.88)
- Ineq-Comp: Benchmarking Human-Intuitive Compositional Reasoning in Automated Theorem Proving of Inequalities (NeurIPS 2025 · similarity 0.87)
- APOLLO: Automated LLM and Lean Collaboration for Advanced Formal Reasoning (NeurIPS 2025 · similarity 0.87)
- Finding Simple Proofs for First-Order Optimization (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
