Can large language models reason about medical questions?

Abstract

Although large language models often produce impressive outputs, it remains unclear how they perform in real-world scenarios requiring strong reasoning skills and expert domain knowledge. We set out to investigate whether closed- and open-source models (GPT-3.5, Llama 2, etc.) can be applied to answer and reason about difficult real-world-based questions. We focus on three popular medical benchmarks (MedQA-US Medical Licensing Examination [USMLE], MedMCQA, and PubMedQA) and multiple prompting scenarios: chain of thought (CoT; think step by step), few shot, and retrieval augmentation. Based on an expert annotation of the generated CoTs, we found that InstructGPT can often read, reason, and recall expert knowledge. Last, by leveraging advances in prompt engineering (few-shot and ensemble methods), we demonstrated that GPT-3.5 not only yields calibrated predictive distributions but also reaches the passing score on three datasets: MedQA-USMLE (60.2%), MedMCQA (62.7%), and PubMedQA (78.2%). Open-source models are closing the gap: Llama 2 70B also passed the MedQA-USMLE with 62.5% accuracy.

Keywords: GPT-3.5; Llama 2; MedQA; large language models; machine learning; medical; open source; prompt engineering; question answering; uncertainty quantification.

Graphical abstract

Figure 1

Answering a USMLE (US Medical Licensing Examination) question using zero-shot CoT prompting “Let’s think step by step” and InstructGPT Selected example.

Figure 2

Prompt templates In the table, we use typewriter style and brackets to represent [provided data] such as the question, additional context, or the answer and generated by GPT-3. The symbol $\varnothing$ represents an empty string.

Figure 3

Generative process and answer likelihood (ensemble model, i.e., self-consistency)

Figure 4

Frequencies of USMLE answers and InstructGPT (text-davinci-002) predictions for direct and CoT prompts (no grounding, zero-shot)

Figure 5

Sampling and combining multiple CoTs Answering accuracy of Codex 5-shot CoT (code-davinci-002) on the USMLE (test), the MedMCQA (validatuin), and the PubMedQA (test) datasets for 100 CoTs sampled with temperature $\tau \in \left\{0,0.5\right\}$. We report the average accuracy for ensemble models evaluated using random subsets of $k^{\prime }=1\dots 100$ CoTs. We report the mean and standard deviation. We display the performances of the best fine-tuned methods along with the lower human baselines.

Figure 6

Uncertainty quantification First row: distribution of the probability assigned to the correct label for correct predictions and incorrect predictions (see Equation 1). Second row: calibration plot. The probabilities are obtained using Codex 5-shot CoT and an ensemble of $k=100$ predictions sampled with temperature $\tau =0.5$.

Figure 7

Comparing open-source LLMs against the closed-source Codex on the MedQA-USMLE benchmark ($\tau =0.9$, up to $k=100$ samples) We report answering accuracy, model calibration, and answering bias.

Figure 8

MedQA-USMLE accuracy vs. model size All experiments were performed using a 5-shot CoT prompting strategy and greedy decoding ($\tau =0$). Llama 2 70B outperforms Codex 175B (proprietary).

Figure 9

(Sample 1) Generated zero-shot CoT from InstructGPT text-davinci-002 for three CoT prompts on a sample for the MedQA-USMLE test set