Research map / Large language models / Agents, code and math
Learning to Prove Theorems by Learning to Generate Theorems
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 10M synthetic theorems from MetaGen-IL to training | Holophrasm · set.mm test, 100% human proofs | theorems proved: 557 -> 600 | Helped |
| add 1M MetaGen-IL synthetic theorems | Holophrasm · set.mm test, 10% human proofs | theorems proved: 454 -> 472 | Helped |
| add 300K synthetic theorems from MetaGen-RL-Adv, no human proofs | Holophrasm · set.mm test, 0% human proofs | theorems proved: 219 -> 357 | Helped |
| add 1M MetaGen-IL synthetic theorems | Holophrasm · iset.mm test | theorems proved: 378 -> 398 | Helped |
| learned generator (IL) instead of random generator (ablation) | Holophrasm · set.mm test, 100% human proofs | theorems proved: 565 -> 574 | Helped |
| learned generator (IL) instead of random generator (ablation) | Holophrasm · set.mm test, 10% human proofs | theorems proved: 457 -> 472 | Helped |
| learned generator (RL-Adv) instead of random generator (ablation) | Holophrasm · set.mm test, 0% human proofs | theorems proved: 346 -> 357 | Helped |
| remove trivial proof steps from relevance network training data (ablation) | Holophrasm + MetaGen-IL · set.mm test, 100% human proofs | theorems proved: 574 -> 600 | Helped |
| add synthetic theorems from MetaGen-Rand (ablation) | Holophrasm · set.mm test, 100% human proofs | theorems proved: 557 -> 565 | Helped |
| add 10M MetaGen-IL synthetic theorems | relevance network · set.mm val, 100% human proofs | Top-1 accuracy: 51.52 -> 53.20 | Helped |
| add 10M MetaGen-IL synthetic theorems | relevance network · set.mm val, 100% human proofs | Top-20 accuracy: 93.41 -> 93.13 | No clear effect |
| add 10M MetaGen-IL synthetic theorems | substitution network · set.mm val, 100% human proofs | accuracy: 81.57 -> 83.90 | Helped |
Closest papers by meaning
- REFACTOR: Learning to Extract Theorems from Proofs (ICLR 2024 · similarity 0.90)
- LEGO-Prover: Neural Theorem Proving with Growing Libraries (ICLR 2024 · similarity 0.88)
- miniCTX: Neural Theorem Proving with (Long-)Contexts (ICLR 2025 · similarity 0.88)
- ProofAug: Efficient Neural Theorem Proving via Fine-grained Proof Structure Analysis (ICML 2025 · similarity 0.88)
- Reviving DSP for Advanced Theorem Proving in the Era of Reasoning Models (NeurIPS 2025 · similarity 0.88)
- Alchemy: Amplifying Theorem-Proving Capability Through Symbolic Mutation (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.