Research map / Large language models / Agents, code and math
Decompose, Structure, and Repair: A Neuro-Symbolic Framework for Autoformalization via Operator Trees
Xiaoyang Liu, Zineng Dong, Yifan Bai, Yantao Li, Yuntian Liu, Tao Luo
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 operator tree supervision to the component formalizer (ablation) | DSR Formalizer (Pass@1) · ProverBench | CC pass rate: 21.54 -> 24.92 | Improved |
| add operator tree supervision without curriculum (ablation) | DSR Formalizer (Pass@1) · PRIME | SC pass rate: 22.44 -> 19.87 | Hurt |
| add curriculum learning on top of operator trees (ablation) | DSR Formalizer (Pass@1) · PRIME | SC pass rate: 19.87 -> 23.08 | Improved |
| add curriculum learning on top of operator trees (ablation) | DSR Formalizer (Pass@1) · ProofNet | CC pass rate: 11.86 -> 16.44 | Improved |
| add operator tree supervision without curriculum (ablation) | DSR Formalizer (Pass@4) · PRIME | CC pass rate: 25.00 -> 21.15 | Hurt |
| add tree-guided repair instead of statement-level global repair (ablation) | DSR · ProofNet | CC pass rate: 76.01 -> 79.51 | Improved |
| add tree-guided repair instead of statement-level global repair (ablation) | DSR · ProverBench | CC pass rate: 82.77 -> 84.00 | Improved |
| add tree-guided repair instead of statement-level global repair (ablation) | DSR · PRIME | SC pass rate: 83.97 -> 80.13 | Hurt |
| add tree-guided repair instead of statement-level global repair (ablation) | DSR · ProofNet | SC pass rate: 89.76 -> 87.33 | Hurt |
| add operator tree supervision to end-to-end statement model (ablation) | End-to-end model (Pass@1) · ProofNet | SC pass rate: 65.77 -> 70.89 | Improved |
| add operator tree supervision to end-to-end statement model (ablation) | End-to-end model (Pass@1) · PRIME | CC pass rate: 44.23 -> 40.38 | Hurt |
| add curriculum learning to end-to-end statement model (ablation) | End-to-end model (Pass@1) · ProofNet | CC pass rate: 41.51 -> 39.08 | No clear effect |
Closest papers by meaning
- Autoformalize Mathematical Statements by Symbolic Equivalence and Semantic Consistency (NeurIPS 2024 · similarity 0.88)
- Neural-Symbolic Recursive Machine for Systematic Generalization (ICLR 2024 · similarity 0.88)
- FormalAlign: Automated Alignment Evaluation for Autoformalization (ICLR 2025 · similarity 0.87)
- ATLAS: Autoformalizing Theorems through Lifting, Augmentation, and Synthesis of Data (NeurIPS 2025 · similarity 0.87)
- Failure by Interference: Language Models Make Balanced Parentheses Errors When Faulty Mechanisms Overshadow Sound Ones (NeurIPS 2025 · similarity 0.87)
- Controlling Equational Reasoning in Large Language Models with Prompt Interventions (AAAI 2025 · similarity 0.86)
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.
