Induction-Deduction Prompting: Enhancing Hidden-Information Reasoning in Medical LLM QA

Anonymous

Abstract: Large Language Models (LLMs) excel in medical question answering. However, limited attention has been given to the underlying reasoning patterns in LLM generated chain of thoughts. We analyse common medical reasoning scenarios using a Bayesian Network, revealing the prevalence of hidden information, especially in the MedQA dataset.We introduce two simple prompts, induction (inferring hidden information) and deduction (evaluating options based on observed and inferred information). Used together they outperform conventional prompting techniques as well as Med-Palm 2, which relies on complex, expert-crafted prompting and expensive fine-tuning.

Paper Type: short

Research Area: Question Answering

Contribution Types: NLP engineering experiment

Languages Studied: English