Research map / Reinforcement learning / Imitation learning
GAR: Generative Adversarial Reinforcement Learning for Formal Theorem Proving
Ruida WANG, Jiarui Yao, Rui Pan, Shizhe Diao, Tong Zhang
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 |
|---|---|---|---|
| apply GAR training | Goedel-Prover-V2-8B · MiniF2F-Test pass@32 | pass rate: 77.87% -> 80.33% | Improved |
| apply GAR training | DeepSeek-Prover-V2-7B · MiniF2F-Test pass@32 | pass rate: 70.49% -> 74.18% | Improved |
| apply GAR training | DeepSeek-Prover-V2-7B · ProofNet-Test | pass rate: 22.58% -> 25.81% | Improved |
| apply GAR training | DeepSeek-Prover-V2-7B · PutnamBench pass@32 | problems solved: 22/660 -> 24/660 | Improved |
| add the statement modification penalty (ablation) | GAR · Statement modification rate, step 4 | modification rate: 74.11% -> 33.63% | Improved |
| use GAR instead of further GRPO on static data (ablation) | Goedel-Prover-V2-8B · MiniF2F-Test pass@32 | pass rate: 77.46% -> 80.33% | Improved |
| add further GRPO on static data (ablation) | Goedel-Prover-V2-8B · MiniF2F-Test pass@32 | pass rate: 77.87% -> 77.46% | No clear effect |
| train the statement fuser too (not frozen) (ablation) | GAR on Goedel-Prover-V2-8B · MiniF2F-Test pass@32 | pass rate: 77.87% -> 80.33% | Improved |
| use problem fusion instead of single-problem MagicCoder evolving (ablation) | GAR on Goedel-Prover-V2-8B · MiniF2F-Test pass@32 | pass rate: 75.41% -> 80.33% | Improved |
| use GAR-trained prover instead of base prover to answer fused statements | statement fuser study · Proof correctness iteration 4 | average proof correctness rate: 7.69% -> 21.79% | Improved |
| apply GAR training | DeepSeek-Prover-V2 · MiniF2F-Test modification rate | statement modification rate: 6.96% -> 13.11% | Hurt |
| apply GAR training | Goedel-Prover-V2 · MiniF2F-Test modification rate | statement modification rate: 24.18% -> 27.05% | Mixed |
Closest papers by meaning
- Generative Adversarial Reasoner: Enhancing LLM Reasoning with Adversarial Reinforcement Learning (ICLR 2026 · similarity 0.88)
- Reviving DSP for Advanced Theorem Proving in the Era of Reasoning Models (NeurIPS 2025 · similarity 0.88)
- DeepSeek-Prover-V1.5: Harnessing Proof Assistant Feedback for Reinforcement Learning and Monte-Carlo Tree Search (ICLR 2025 · similarity 0.88)
- STP: Self-play LLM Theorem Provers with Iterative Conjecturing and Proving (ICML 2025 · similarity 0.87)
- LEGO-Prover: Neural Theorem Proving with Growing Libraries (ICLR 2024 · similarity 0.87)
- Subgoal-based Demonstration Learning for Formal Theorem Proving (ICML 2024 · similarity 0.86)
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.
