Currently, large model research is gradually shifting from scaling law–driven pre-training to post-training that focuses on reasoning capabilities. Given the effectiveness and universality of symbolic logic reasoning, enhancing the logical reasoning ability of large models has become a key approach to addressing the issue of hallucination.
To advance research on the logical reasoning capabilities of large language models, Tsinghua University – University of Amsterdam Joint Research Centre for Logic, in collaboration with researchers from four other universities—including Peking University, the University of Amsterdam (UvA), Carnegie Mellon University (CMU), and Mohamed bin Zayed University of Artificial Intelligence (MBZUAI)—recently published a paper titled "Empowering LLMs with Logical Reasoning: A Comprehensive Survey."
The paper provides a comprehensive survey of cutting-edge research methods and evaluation benchmarks in the field. Centered around two key scientific questions—logical question answering and logical consistency—it establishes a complete classification framework and systematically summarizes state-of-the-art approaches. Additionally, it compiles commonly used public benchmark datasets and evaluation metrics in the field, and explores important future research directions worthy of attention.
Research Survey Classification Framework as proposed in the paper
The paper has been accepted in the Survey Track of IJCAI 2025, one of the most prestigious conferences in the field of artificial intelligence, organized by the Association for the Advancement of Artificial Intelligence. The author team will also deliver a tutorial on the same topic at the conference, providing a comprehensive discussion on the challenges, mainstream approaches, and future opportunities in this research area. IJCAI is one of the most influential international conferences in AI and currently the only top-tier conference with a dedicated track for survey papers.
Professor Liu Fenrong from the School of Humanities at Tsinghua University and her student Cheng Fengxiang contributed to this work. Cheng Fengxiang received her master’s degree from Tsinghua University in 2024 and is currently pursuing a Ph.D. at the University of Amsterdam under the supervision of Professor Liu Fenrong and Professor van Rooij from the University of Amsterdam. The project also involved Li Haoxuan, Assistant Researcher at Peking University; Professor Lin Zhouchen from Peking University; Professor Zhang Kun from Carnegie Mellon University and Mohamed bin Zayed University of Artificial Intelligence; and Professor van Rooij.
Full paper can be accessed at the following link:
https://arxiv.org/abs/2502.15652
Editor：Li Han
