Disentangling Reasoning and Knowledge in Medical Large Language Models

TL;DR

This work confronts the challenge that medical reasoning benchmarks conflate reasoning with factual recall. It introduces a PubMedBERT-based labeling system to separate reasoning-heavy from knowledge-heavy questions across 11 biomedical QA benchmarks, revealing that only about one-third require complex reasoning. Through systematic experiments with biomedical and general-domain LLMs, the authors show a persistent gap between knowledge and reasoning performance and demonstrate that robustness to adversarial prompts is limited, especially for reasoning. They introduce BioMed-R1, trained with supervised fine-tuning on reasoning-heavy data and reinforced through RL, achieving strong performance for its size and improving adversarial robustness; results also highlight the value of combining reasoning supervision with RL and adversarial traces. The findings motivate richer reasoning-focused data (e.g., clinical case reports) and adaptive generation strategies to advance safe, reliable biomedical reasoning in LLMs.

Abstract

Medical reasoning in large language models (LLMs) aims to emulate clinicians' diagnostic thinking, but current benchmarks such as MedQA-USMLE, MedMCQA, and PubMedQA often mix reasoning with factual recall. We address this by separating 11 biomedical QA benchmarks into reasoning- and knowledge-focused subsets using a PubMedBERT classifier that reaches 81 percent accuracy, comparable to human performance. Our analysis shows that only 32.8 percent of questions require complex reasoning. We evaluate biomedical models (HuatuoGPT-o1, MedReason, m1) and general-domain models (DeepSeek-R1, o4-mini, Qwen3), finding consistent gaps between knowledge and reasoning performance. For example, HuatuoGPT-o1 scores 56.9 on knowledge but only 44.8 on reasoning. In adversarial tests where models are misled with incorrect initial reasoning, biomedical models degrade sharply, while larger or RL-trained general models show more robustness. To address this, we train BioMed-R1 using fine-tuning and reinforcement learning on reasoning-heavy examples. It achieves the strongest performance among similarly sized models. Further gains may come from incorporating clinical case reports and training with adversarial and backtracking scenarios.

Figures (19)

• Figure 1: Representative examples of knowledge-heavy and reasoning-heavy questions, sampled from benchmark datasets. Knowledge-heavy questions tend to be shorter and require direct factual recall, whereas reasoning-heavy questions are typically longer, involving multi-step inference.
• Figure 2: Overview of our evaluation pipeline. (A) A PubMedBERT classifier is trained on MedXpert to distinguish between knowledge- and reasoning-heavy questions. (B) The classifier achieves 81% agreement with both gold-standard labels and expert annotations. (C) We apply it across benchmarks to stratify question types, revealing consistent performance gaps. Finally, we test robustness under adversarial prompts and find most models struggle to recover.
• Figure 3: Accuracy of the PubMedBERT classifier, a medical expert, and a word-count-based logistic regression baseline on the MedXpert and out-of-distribution (OOD) datasets. PubMedBERT outperforms the baseline and matches expert-level accuracy (81% on MedXpert, 85% on OOD), demonstrating strong generalization and alignment with both expert and gold-standard labels.
• Figure 4: Average token counts generated by models trained with SFT and SFT + RL, stratified by reasoning-heavy and knowledge-heavy questions across all evaluation benchmarks. Models trained with RL on top of SFT tend to produce more concise reasoning traces compared to SFT-only models, despite both being trained on distilled traces from DeepSeek-R1. As expected, all models generate longer outputs for reasoning-heavy questions, reflecting their higher difficulty.
• Figure S1: Agreement between MedXpert labels and an independent set of 100 samples annotated by a separate medical expert. Although the original labels were also expert-generated, the 79% agreement underscores the task’s subjectivity and the potential for inter-annotator variability.