Research map / Architectures and efficiency / Neural architecture search
HyperTree Proof Search for Neural Theorem Proving
Guillaume Lample, Timothee Lacroix, Marie-anne Lachaux, Aurelien Rodriguez, Amaury Hayat, Thibaut Lavril, Gabriel Ebner, Xavier Martinet
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 online training on proof-search data to supervised model | Evariste on Metamath · Metamath test pass@32 | pass rate %: 61.2% -> 72.4% | Improved |
| add online training on unproved statements | Evariste on Equations Identities · Equations Identities | cumulative pass rate %: never exceeds 36% -> 91.3% | Improved |
| train policy on minimal proofs instead of all proofs of solved nodes (ablation) | Evariste · Equations Identities | cumulative pass rate %: 40.6 -> 78.1 | Improved |
| train policy on root minimal proof instead of all nodes (ablation) | Evariste · Metamath valid | pass@8 %: 51.6 -> 68.6 | Improved |
| restrict to minimal proofs on Metamath test (ablation) | Evariste · Metamath test | pass@8 %: 57.2 -> 58.8 | Improved |
| add critic model (ablation) | Evariste · Equations Identities | cumulative pass rate %: 65.6 -> 78.1 | Improved |
| add critic model (ablation) | Evariste · Metamath valid | pass@8 %: 64.8 -> 68.6 | Improved |
| soft critic targets (HTPS action values) instead of hard 0/1 targets (ablation) | Evariste · Equations Identities | cumulative pass rate %: 63.1 -> 78.1 | Improved |
| sample search hyper-parameters per attempt instead of fixed (ablation) | Evariste · Equations Identities | cumulative pass rate %: 73.8 -> 78.1 | Improved |
| sample search hyper-parameters per attempt instead of fixed (ablation) | Evariste · Metamath valid | pass@8 %: 69.8 -> 68.6 valid; 56.2 -> 57.4 test | Mixed |
| add synthetic Equations co-training task (ablation) | Evariste on Lean · miniF2F-valid cumulative | cumulative pass rate %: 54.9% -> 58.6% | Improved |
| refresh model on provers more frequently (online vs expert iteration) (ablation) | Evariste · Equations/Metamath | cumulative pass rate | Improved |
Closest papers by meaning
- LEGO-Prover: Neural Theorem Proving with Growing Libraries (ICLR 2024 · similarity 0.90)
- ProofAug: Efficient Neural Theorem Proving via Fine-grained Proof Structure Analysis (ICML 2025 · similarity 0.89)
- REFACTOR: Learning to Extract Theorems from Proofs (ICLR 2024 · similarity 0.89)
- miniCTX: Neural Theorem Proving with (Long-)Contexts (ICLR 2025 · similarity 0.88)
- Alchemy: Amplifying Theorem-Proving Capability Through Symbolic Mutation (ICLR 2025 · similarity 0.88)
- Reviving DSP for Advanced Theorem Proving in the Era of Reasoning Models (NeurIPS 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.
