Research map / Graph learning / Knowledge and temporal graphs
ProofFlow: A Dependency Graph Approach to Faithful Proof Autoformalization
Rafael Medeiros Cabral, Tuan Manh Do, Yu Xuejun, Wai Ming Tai, Zijin Feng, SHEN XIN
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 |
|---|---|---|---|
| use lemma-based DAG pipeline instead of full-proof formalization | LLM autoformalization · ProofFlowBench Pass@5 (thinking mode) | ProofScore: 0.123 -> 0.545 | Improved |
| use lemma-based DAG pipeline instead of step-proof formalization | LLM autoformalization · ProofFlowBench Pass@5 (thinking mode) | ProofScore: 0.072 -> 0.545 | Improved |
| enforce dependency DAG (ablation) | ProofFlow · ProofFlowBench Pass@5 (thinking mode) | ProofScore: 0.417 -> 0.545 | Improved |
| enforce dependency DAG (ablation) | ProofFlow · ProofFlowBench Pass@5 (non-thinking mode) | ProofScore: 0.347 -> 0.355 | No clear effect |
| enforce dependency DAG (ablation) | ProofFlow · ProofFlowBench Pass@5 (thinking mode) | proof-level correct syntax: 0.353 -> 0.375 | Improved |
| enforce dependency DAG (ablation) | ProofFlow · ProofFlowBench Pass@5 (non-thinking mode) | proof-level correct syntax: 0.049 -> 0.027 | Hurt |
| enforce dependency DAG (ablation) | ProofFlow · ProofFlowBench Pass@5 (non-thinking mode) | formalizer accuracy: 0.807 -> 0.751 | Hurt |
| enforce dependency DAG (ablation) | ProofFlow · ProofFlowBench Pass@5 (thinking mode) | tactic accuracy: 0.681 -> 0.742 | Improved |
| enforce dependency DAG (ablation) | ProofFlow · Error analysis, thinking mode | error-free steps (%): 42.8 -> 53.3 | Improved |
| enforce dependency DAG (ablation) | ProofFlow · ProofFlowBench Pass@5 (thinking mode) | output tokens (k): 98.5 -> 94.2 | Improved |
| enable thinking mode (stronger models) | ProofFlow DAG · ProofFlowBench Pass@5 | ProofScore: 0.355 -> 0.545 | Improved |
| increase retries from Pass@1 to Pass@5 (ablation) | ProofFlow DAG · ProofFlowBench Pass@5 (thinking mode) | ProofScore: 0.508 -> 0.545 | Improved |
Closest papers by meaning
- FaithSieve: Fine-Grained Evaluation of Math Proofs with Faithful Formal Evidence (NeurIPS 2026 · similarity 0.90)
- APE-Bench: Evaluating Automated Proof Engineering for Formal Math Libraries (ICML 2026 · similarity 0.89)
- APOLLO: Automated LLM and Lean Collaboration for Advanced Formal Reasoning (NeurIPS 2025 · similarity 0.89)
- ImProver: Agent-Based Automated Proof Optimization (ICLR 2025 · similarity 0.89)
- FormalAlign: Automated Alignment Evaluation for Autoformalization (ICLR 2025 · similarity 0.89)
- Not All Proofs Are Equal: Evaluating LLM Proof Quality Beyond Correctness (NeurIPS 2026 · 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.
