DiagX-DT: diagnostic exclusion reasoning with dialectical thinking for traditional Chinese medicine in large language models

The essence of syndrome differentiation in Traditional Chinese Medicine (TCM) lies in dynamically filtering and associating symptoms to reach dialectical conclusions. However, while large language models (LLMs) have shown promise in various domains, they struggle to replicate the context-sensitive and exclusionary reasoning required for syndrome differentiation in TCM, primarily because they are typically trained on structured data that does not account for the complex and iterative reasoning process inherent in TCM. In this paper, we propose DiagX-DT, a novel diagnostic framework that integrates dialectical thinking with exclusion-based reasoning to enhance TCM diagnosis using LLMs. Our method first prompts the LLM to generate an initial answer along with a multi-step reasoning chain, while simultaneously extracting the initial probability distribution over candidate options from the model’s final output layer. Then, external structured TCM knowledge bases are utilized to iteratively exclude implausible choices based on treatment contradictions and diagnostic mismatches. After each exclusion, the remaining option set is used to recompute a normalized probability distribution. A recombination-evaluation module dynamically selects the top-scoring candidate under a confidence threshold mechanism to produce the final diagnosis. We evaluate DiagX-DT on three benchmark TCM choice diagnostic datasets, and experimental results demonstrate that our approach significantly improves diagnostic accuracy and robustness compared to standard LLM baselines.