一句话:在 Isabelle/HOL 上比较数学 LLM、推理 LLM 及混合模型在基座、微调与 RAG 三种配置下的证明推断能力。
软件形式验证依赖 Isabelle/HOL 等证明助手来形式化并验证数学定理,但证明构造通常需要人类专家。本文研究大语言模型用于 Isabelle/HOL 证明推断的有效性,系统比较数学 LLM、推理型 LLM 以及混合数学-推理 LLM,并在三种配置下评估:未经领域特定训练的基座模型、面向形式证明领域微调的变体,以及基于检索增强生成(RAG)的系统。该比较不仅具有理论意义,也有望指导利用各类方法互补优势的混合系统开发,从而推动形式验证中自动证明生成的研究。
Software formal verification is an important and challenging tool to ensure software correctness. It requires tools to formulate and verify mathematical theorems. One example of such a tool is Isabelle/HOL, a proof assistant that supports higher-order logic (HOL), enabling rigorous mathematical reasoning that could be useful for formal verification in software and hardware systems. However, the use of these tools requires a human expert on theorem proving in many cases. With the advent of LLMs (Large Language Models) a research line emerged to use them for proof inference. This paper investigates the effectiveness of math LLMs, reasoner LLMs, and hybrid math-reasoner LLMs in Isabelle/HOL proof inference, evaluating their capabilities across three configurations: base models without domain-specific training, fine-tuned variants adapted to the formal proof domain, and systems based on Retrieval Augmented Generation (RAG). The significance of this comparison extends beyond theoretical interest, potentially informing the development of hybrid systems that leverage the complementary strengths of all approaches.