Abstract
Translating natural language (NL) mathematical proofs into formal languages (FL) like Lean 4 remains a significant challenge due to the "formalization gap." ProofBridge addresses this by introducing an end-to-end framework that combines cross-modal retrieval in a joint embedding space with an iterative repair loop. By representing Lean proofs as Directed Acyclic Graphs (DAGs) of tactics and states, ProofBridge captures logical structure more effectively than text-based models. Experimental results on the MINIF2F-TEST-PF benchmark show a 31.14% improvement in semantic correctness over existing state-of-the-art models.
Introduction
Mathematical reasoning in humans is primarily expressed through natural language, which is often intuitive but can be ambiguous or omit rigorous steps. Conversely, formal verification systems like Lean 4 require precise, step-by-step logic. Bridging this gap has traditionally required human experts to manually translate theorems and proof sketches. ProofBridge aims to automate this "holy grail" of auto-formalization: taking a human-written theorem and its proof and producing a complete, verifiable Lean 4 equivalent.
Methodology
The ProofBridge framework rests on three primary pillars: Joint Semantic Embedding, Retrieval-Augmented Fine-Tuning (RAFT), and Iterative Proof Repair.
Joint Semantic Embedding and Retrieval
To facilitate translation, ProofBridge maps NL theorem-proof pairs and FL (Lean) pairs into a shared mathematical space. This allows for cross-modal retrieval, where a new English proof can be matched with semantically similar Lean proofs from a database. These retrieved examples serve as in-context templates for the LLM. Crucially, the system encodes Lean proofs as Directed Acyclic Graphs (DAGs) rather than plain text to better capture the logic of state transformations.
Structural Representation via DAGs
A core innovation of ProofBridge is the treatment of Lean proofs as sequences of state transformations. By encoding proofs as DAGs of tactics and states, the model captures the underlying mathematical logic more robustly than standard text encoders.
Iterative Proof Repair
To mitigate hallucinations and syntax errors, ProofBridge employs a feedback loop. Generated Lean code is sent to the Lean compiler; if it fails, the error message is fed back into the LLM for correction. Additionally, an "LLM judge" (such as Gemini-2.5-Pro) verifies semantic equivalence between the original NL theorem and the generated FL theorem by attempting to prove a bi-directional equivalence within Lean itself.
Evaluation and Results
The system was evaluated using the newly curated NUMINAMATH-LEAN-PF dataset (38.9k pairs) and the MINIF2F-TEST-PF benchmark.
- Retrieval Performance: The DAG-aware joint embedding model achieved a 3.28x higher Recall@1 compared to standard baseline encoders, validating the structural approach.
- End-to-End Accuracy: On MINIF2F-TEST-PF, ProofBridge demonstrated a +31.14% increase in Semantic Correctness and a +1.64% increase in Type Correctness (at pass@32) over the Kimina-Prover baseline, outperforming heavyweights like GPT-5-mini and DeepSeek-Prover.
- Domain Strengths: The framework is particularly effective in Number Theory, solving over 85% of problems, though high-level competition math (e.g., IMO) remains a challenge, with a ~35% success rate.
Conclusion
ProofBridge represents a significant advancement in the auto-formalization of mathematical proofs. By integrating structural DAG-based embeddings with iterative verification and repair, it narrows the gap between human-readable math and machine-verifiable logic. Future work will focus on scaling the repair loop and enhancing the system's ability to handle creative leaps required for Olympiad-level problems.
Summary updated