Do "New Snow Tablets" Contain Snow? Large Language Models Over-Rely on Names to Identify Ingredients of Chinese Drugs

TL;DR

This work evaluates general and TCM-specialized LLMs on identifying ingredients of Traditional Chinese Medicine drugs, exposing systemic reliance on drug names over pharmacological knowledge. It introduces two evaluation tasks (direct ingredient inquiry and ingredient list verification) and identifies three reproducible failure patterns: literal interpretation, overreliance on common herbs, and erratic behavior. A Retrieval Augmented Generation (RAG) approach using the Pharmacopoeia with a Chinese baseline model substantially improves verification accuracy from about $50\%$ to $82\%$, demonstrating a practical path to improved clinical reliability. The findings underscore fundamental knowledge gaps in current domain-specific LLMs and offer a scalable augmentation strategy to bridge theory and practice in TCM ingredient recognition.

Abstract

Traditional Chinese Medicine (TCM) has seen increasing adoption in healthcare, with specialized Large Language Models (LLMs) emerging to support clinical applications. A fundamental requirement for these models is accurate identification of TCM drug ingredients. In this paper, we evaluate how general and TCM-specialized LLMs perform when identifying ingredients of Chinese drugs. Our systematic analysis reveals consistent failure patterns: models often interpret drug names literally, overuse common herbs regardless of relevance, and exhibit erratic behaviors when faced with unfamiliar formulations. LLMs also fail to understand the verification task. These findings demonstrate that current LLMs rely primarily on drug names rather than possessing systematic pharmacological knowledge. To address these limitations, we propose a Retrieval Augmented Generation (RAG) approach focused on ingredient names. Experiments across 220 TCM formulations show our method significantly improves accuracy from approximately 50% to 82% in ingredient verification tasks. Our work highlights critical weaknesses in current TCM-specific LLMs and offers a practical solution for enhancing their clinical reliability.

Figures (16)

• Figure 1: The change of the market size of Chinese proprietary drugs, traditional Chinese drugs and overall traditional Chinese medicine from 2019 to 2024.
• Figure 2: LLMs tend to misunderstand a Chinese drug ingredient by its name because of the literal information is ambiguous.
• Figure 3: In the left figure, we calculate the frequency of all ingredients included in our dataset. The x-axis value represents the oracle truth frequency according to our dataset and the y-axis value represents the frequency of the ingredient appearing in the responses of the TCM-specific LLM BianCang-Qwen2.5-7B-Instruct. In the right figures, we calculate the precision and recall scores of the 10 ingredients with the highest oracle truth frequencies ( upper) and the 10 ingredients with the lowest oracle truth frequencies ( lower).
• Figure 4: The accuracy score of each LLM is calculated on questions expected to get the answer "No" and "Yes", respectively. The bias between them reveals that language models tend to answer with the same binary decision whatever the question is.
• Figure 5: Erratic behavior of GPT-3.5-Turbo.