ExplainCPE: A Free-text Explanation Benchmark of Chinese Pharmacist Examination

Language bias in AI evaluation is a real problem: Existing explanation datasets are predominantly in English and focused on general knowledge, meaning we lack adequate benchmarks to test how well AI systems can explain their reasoning in other languages like Chinese and in specialized domains like medicine.

Current LLMs struggle with medical reasoning and text understanding: When tested on the Chinese Pharmacist Examination benchmark, both ChatGPT and GPT-4 showed limitations in understanding complex medical text and performing computational reasoning—suggesting these systems may not be ready for high-stakes applications without further improvement.

Explainability matters as much as accuracy: The paper emphasizes that it's not enough for AI to give correct answers; the ability to generate clear, accurate reasons for those answers is crucial for trustworthiness, especially in medical contexts where patients and professionals need to understand the reasoning behind recommendations.

Different AI models learn differently: The research found that different LLMs have varying preferences for how they learn from examples (in-context learning), meaning a one-size-fits-all approach to AI evaluation and deployment won't work across different systems.

Medical AI safety requires specialized testing: This work highlights that AI safety and trustworthiness in healthcare specifically need dedicated research and benchmarks, not just general-purpose AI evaluation—a critical gap for responsible deployment in medical settings.