Research map / Continual learning / Domain and prompt-based
OptProver: Bridging Olympiad and Optimization through Continual Training in Formal Theorem Proving
Chenyi Li, Yanchen Nie, Zhenyu Ming, Gong Zhang, Kun Yuan, Zaiwen Wen
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 two rounds of expert iteration to BFS-Prover-V2 | BFS-Prover-V2-7B · OptBench | Pass@1 average: 13.25% -> 28.75% | Improved |
| add two rounds of expert iteration to BFS-Prover-V2 | BFS-Prover-V2-7B · OptBench | Pass@32 average: 32.00% -> 50.00% | Improved |
| add DPO after expert iteration (ablation) | OptProver (Only EI) · OptBench | Pass@32 average: 50.00% -> 52.00% | Improved |
| use UAPO instead of DPO (ablation) | OptProver (EI) · OptBench | Pass@32 average: 52.00% -> 53.50% | Improved |
| use PW-DPO instead of DPO (ablation) | OptProver (EI) · OptBench | Pass@32 average: 52.00% -> 54.25% | Improved |
| combine utility-aware and perplexity-weighted (PW-UAPO) (ablation) | OptProver (EI) · OptBench | Pass@32 average: 54.25% -> 55.25% | Improved |
| use step-level proving instead of whole-proof generation | 7B provers · OptBench | average Pass@32: 19.75% -> 32.00% | Improved |
| set PW-DPO tau to median 1.08 instead of 1.00 (ablation) | PW-DPO · OptBench | Pass@1: 36.00% -> 36.25% | Improved |
| raise PW-DPO tau to 2.42 (ablation) | PW-DPO · OptBench | Pass@1: 36.25% -> 32.75% | Hurt |
| apply OptProver training to base model | BFS-Prover-V2-7B · ProofNet | problems solved at Pass@1: 21 -> 31 | Improved |
| apply OptProver training to base model | BFS-Prover-V2-7B · ProofNet | problems solved at Pass@32: 42 -> 49 | Improved |
| apply OptProver training to base model | BFS-Prover-V2-7B · MiniF2F | Pass@32 | No clear effect |
Closest papers by meaning
- ImProver: Agent-Based Automated Proof Optimization (ICLR 2025 · similarity 0.89)
- Finding Simple Proofs for First-Order Optimization (NeurIPS 2026 · similarity 0.88)
- Subgoal-based Demonstration Learning for Formal Theorem Proving (ICML 2024 · similarity 0.87)
- MPS-Prover: Advancing Stepwise Theorem Proving by Multi-Perspective Search and Data Curation (NeurIPS 2025 · similarity 0.87)
- BAIT: Benchmarking (Embedding) Architectures for Interactive Theorem-Proving (AAAI 2024 · similarity 0.87)
- LeanAgent: Lifelong Learning for Formal Theorem Proving (ICLR 2025 · similarity 0.87)
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.
