Do Large Language Models Excel in Complex Logical Reasoning with Formal Language?
This work addresses the need for systematic evaluation of LLMs in logical reasoning with formal languages, providing insights for researchers and practitioners, though it is incremental in nature.
The paper conducted a comprehensive evaluation of large language models (LLMs) on complex logical reasoning tasks using formal languages, finding that thinking models outperform instruct models, LLMs have limitations in inductive reasoning, and data with PoT format yields the best generalization, with a simple fine-tuning method improving performance.
Large Language Models (LLMs) have been shown to achieve breakthrough performance on complex logical reasoning tasks. Nevertheless, most existing research focuses on employing formal language to guide LLMs to derive reliable reasoning paths, while systematic evaluations of these capabilities are still limited. In this paper, we aim to conduct a comprehensive evaluation of LLMs across various logical reasoning problems utilizing formal languages. From the perspective of three dimensions, i.e., spectrum of LLMs, taxonomy of tasks, and format of trajectories, our key findings are: 1) Thinking models significantly outperform Instruct models, especially when formal language is employed; 2) All LLMs exhibit limitations in inductive reasoning capability, irrespective of whether they use a formal language; 3) Data with PoT format achieves the best generalization performance across other languages. Additionally, we also curate the formal-relative training data to further enhance the small language models, and the experimental results indicate that a simple rejected fine-tuning method can better enable LLMs to generalize across formal languages and achieve the best overall performance. Our codes and reports are available at https://github.com/jiangjin1999/FormalEval.