F2: Let Large Language Models Think like Aristotle

ICLR 2026 Conference Submission16656 Authors

Keywords: Hallucination Detection, Logical Inference, Large Language Models

Abstract: With the wide application of Large Language Models (LLMs), the accuracy and reliability of the content they generate have become the focus of attention. The hallucination of the content generated by the large language model seriously affects the credibility and practicability of the model in key scenarios. However, the mainstream hallucination detection technology relies on external knowledge bases to verify the authenticity of the content generated by the model, or uses a large number of annotation data for training. These methods require complex model structure and support, will consume a lot of resources and time, and cross domain generalization ability is poor. In this paper, we proposes a new hallucination detection method, which allows the LLMs to imitate the way of thinking of the philosopher Aristotle. We decompose the complex hallucination verification process into two distinct subjects (called F2): a. Factual hallucination detection: verifying the fact finding of the content generated by the model and analyzing the optimal solution; b. Fidelity hallucination detection: Logical verification based on classical logical forms, including the reasoning systems or logical forms of logic such as Aristotle's most outstanding contribution, syllogism. The experimental results show that this method not only improves the recognition and analysis ability of the LLMs itself for illusory content, but also enhances the interpretability of the defects of the LLMs, enabling the developers of the LLMs to effectively identify the sources of errors and improve the model capabilities.

Primary Area: neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)

Submission Number: 16656