Hierarchical Attention Generates Better Proofs

2025-04-27

AI Summary

Large language models (LLMs) struggle to model the hierarchical structure of mathematical reasoning in formal theorem proving. Method: We propose Hierarchical Attention Regularization (HAR), the first approach to explicitly encode a five-level mathematical reasoning hierarchy—proposition → lemma → step → subgoal → atomic operation—into the Transformer attention mechanism via structural alignment constraints. HAR integrates mathematically grounded attention regularization with formal proof fine-tuning on miniF2F and ProofNet. Contribution/Results: On miniF2F, HAR improves proof success rate by 2.05% and reduces average proof steps by 23.81%; on ProofNet, it achieves a 1.69% gain in success rate and a 16.50% reduction in proof steps. These results demonstrate substantially enhanced modeling of abstract reasoning structures and improved generation efficiency for formal proofs.

Technology Category

Knowledge Representation and Reasoning: Automated Reasoning and Theorem ProvingNatural Language Processing: Sentence-level Semantics, Textual Inference, etc.Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning

Application Category

Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMsSearch and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for rankingGraph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Large language models (LLMs) have shown promise in formal theorem proving, but their token-level processing often fails to capture the inherent hierarchical nature of mathematical proofs. We introduce extbf{Hierarchical Attention}, a regularization method that aligns LLMs' attention mechanisms with mathematical reasoning structures. Our approach establishes a five-level hierarchy from foundational elements to high-level concepts, ensuring structured information flow in proof generation. Experiments demonstrate that our method improves proof success rates by 2.05% on miniF2F and 1.69% on ProofNet while reducing proof complexity by 23.81% and 16.50% respectively. The code is available at https://github.com/Car-pe/HAGBP.

Problem

• Aligns LLM attention with mathematical proof structures
• Improves proof success rates on miniF2F and ProofNet
• Reduces proof complexity by significant percentages

Innovation

• Hierarchical Attention aligns LLMs with math structures
• Five-level hierarchy ensures structured proof generation
• Improves success rates and reduces proof complexity