Research map / Large language models / Language, safety and interpretability
ProofBridge: Auto-Formalization of Natural Language Proofs in Lean via Joint Embeddings
Prithwish Jana, Kaan Kale, Ahmet Ege Tanriverdi, Cruise Song, Sriram Vishwanath, Vijay Ganesh
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 retrieved relevant FL examples to SFT input (retrieval-augmented SFT) (ablation) | ProofBridge (Kimina-Prover-RL-1.7B) · miniF2F-Test-PF | Semantic correctness pass@32 (%): 34.84 -> 55.33 | Improved |
| add retrieved relevant FL examples to SFT input (retrieval-augmented SFT) (ablation) | ProofBridge (Kimina-Prover-RL-1.7B) · miniF2F-Test-PF | Type correctness pass@32 (%): 78.69 -> 89.75 | Improved |
| add iterative proof repair with Lean feedback and equivalence feedback (ablation) | ProofBridge (Retrieval-augmented SFT) · miniF2F-Test-PF | Semantic correctness pass@32 (%): 55.33 -> 62.70 | Improved |
| add iterative proof repair with Lean feedback and equivalence feedback (ablation) | ProofBridge (Retrieval-augmented SFT) · miniF2F-Test-PF | Type correctness pass@32 (%): 89.75 -> 95.49 | Improved |
| fine-tune with SFT on NL-FL pairs, evaluated with joint-embedding few-shot examples (ablation) | Kimina-Prover-RL-1.7B (text-based retrieval few-shot) · miniF2F-Test-PF | SC / TC (relative change): +2.06% SC, -11.06% TC | Mixed |
| add five random few-shot examples | Kimina-Prover-RL-1.7B · miniF2F-Test-PF | Semantic correctness pass@32 (%): 40.16 -> 31.56 | Hurt |
| add five random few-shot examples | Kimina-Prover-RL-1.7B · miniF2F-Test-PF | Type correctness pass@32 (%): 75.00 -> 93.85 | Improved |
| add text-based retrieved few-shot examples (Qwen3-Embedding-8B) vs random few-shot | Kimina-Prover-RL-1.7B · miniF2F-Test-PF | SC / TC pass@32 (%): SC 31.56 -> 32.78; TC 93.85 -> 89.75 | Mixed |
| replace baseline NL encoder with contrastively trained joint NL-FL embedding | NL->FL retrieval · NuminaMath-Lean-PF test | Recall@1 (%): 16.06 -> 52.83 | Improved |
| use joint embedding instead of larger SoTA encoder Qwen3-Embedding-8B | NL->FL retrieval · NuminaMath-Lean-PF test | Recall@1 (%): 46.75 -> 52.83 | Improved |
| two-step pipeline (theorem autoformalizer then prover) vs end-to-end | Herald-Translator -> Kimina-Prover-Distill-8B · miniF2F-Test-PF | Type correctness pass@32 (%): 59.43 | Hurt |
Closest papers by meaning
- FaithSieve: Fine-Grained Evaluation of Math Proofs with Faithful Formal Evidence (NeurIPS 2026 · similarity 0.90)
- ProofAug: Efficient Neural Theorem Proving via Fine-grained Proof Structure Analysis (ICML 2025 · similarity 0.90)
- ImProver: Agent-Based Automated Proof Optimization (ICLR 2025 · similarity 0.89)
- ATLAS: Autoformalizing Theorems through Lifting, Augmentation, and Synthesis of Data (NeurIPS 2025 · similarity 0.89)
- APOLLO: Automated LLM and Lean Collaboration for Advanced Formal Reasoning (NeurIPS 2025 · similarity 0.89)
- Lean-STaR: Learning to Interleave Thinking and Proving (ICLR 2025 · similarity 0.89)
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.
