Disentangling Reasoning and Knowledge in Medical Large Language Models

2025-05-16

AI Summary

Existing medical LLM evaluation benchmarks (e.g., MedQA-USMLE) conflate reasoning capability with factual memorization, hindering accurate assessment of clinical diagnostic reasoning. Method: This work systematically disentangles reasoning and knowledge components in medical question answering for the first time, constructing the first dual-track benchmark subset—classified via PubMedBERT (81% accuracy)—to isolate reasoning-demanding items (only 32.8% require complex reasoning). We propose an adversarial misdirection detection and backtracking reasoning training paradigm, integrating supervised fine-tuning and reinforcement learning (RL) on clinical case–enhanced data to train BioMed-R1. Contribution/Results: BioMed-R1 achieves state-of-the-art reasoning performance among models of comparable scale; its m1 variant exhibits a 13.4-point gap between knowledge and reasoning scores, highlighting reasoning-specific gains. RL training significantly improves reasoning accuracy, and applying this paradigm to general-purpose LLMs markedly enhances their reasoning robustness.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Knowledge Representation and Reasoning: Diagnosis and Abductive Reasoning
• Natural Language Processing: Safety and Robustness

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

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, m1 scores 60.5 on knowledge but only 47.1 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.

Problem

• Separating reasoning and knowledge in medical LLM benchmarks
• Evaluating performance gaps in biomedical vs general models
• Improving robustness in adversarial medical reasoning tasks

Innovation

• Separates reasoning and knowledge using PubMedBERT classifier
• Trains BioMed-R1 with fine-tuning and reinforcement learning
• Incorporates clinical case reports and adversarial scenarios