Bridging the Gap Between Fuzzy Logic and Large Language Models

Connectionism and symbolism are two influential traditions in artificial intelligence. Large language models (LLMs) exemplify connectionist learning, while fuzzy logic expresses numerical reasoning through graded linguistic rules. We introduce Fuzzy-LLM, a framework that connects these traditions by using a pretrained LLM to construct an interpretable fuzzy classifier from labeled numerical examples without parameter fine-tuning. After feature normalization, the LLM describes values using five predefined linguistic terms and generates IF–THEN rules. For rules with identical antecedents but different class consequents, hard correction selects one class, whereas soft correction retains candidate classes with graded certainty. The rules are translated into a Mamdani inference system that makes predictions without further LLM calls. For each prediction, membership degrees, activated rules, and class scores trace the decision from numerical input to output. Across Iris, Ecoli, Glass, Seeds, and UKM, the best observed accuracies among the evaluated configurations are 94.67%, 82.14%, 61.68%, 94.29%, and 76.73%, respectively. Fuzzy-LLM exceeds the recorded direct-LLM baselines on all five datasets. This combination of language-guided rule construction and explicit fuzzy inference supports inspectable tabular classification.