Automating Expert-Level Medical Reasoning Evaluation of Large Language Models

2025-07-10

AI Summary

Existing evaluation methods for medical reasoning in large language models (LLMs) lack expert-level rigor, transparency, and scalability. Method: We introduce MedThink-Bench—the first benchmark comprising 500 high-difficulty medical questions, each annotated with expert-curated, fine-grained reasoning chains—and propose LLMS-w-Ref, an evaluation framework integrating reasoning-chain supervision with an LLM-as-a-Judge mechanism to ensure both scalability and human-expert alignment. Contribution/Results: Comprehensive evaluation of 12 state-of-the-art models reveals that smaller open-source models—e.g., MedGemma-27B—can outperform larger proprietary models on medical reasoning tasks. MedThink-Bench achieves 92.3% agreement with human experts, validating its effectiveness, interpretability, and practical utility for diagnostic reasoning assessment.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Knowledge Representation and Reasoning: Computational Complexity of Reasoning
• Reasoning under Uncertainty: Other Foundations of Reasoning under Uncertainty

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

As large language models (LLMs) become increasingly integrated into clinical decision-making, ensuring transparent and trustworthy reasoning is essential. However, existing evaluation strategies of LLMs' medical reasoning capability either suffer from unsatisfactory assessment or poor scalability, and a rigorous benchmark remains lacking. To address this, we introduce MedThink-Bench, a benchmark designed for rigorous, explainable, and scalable assessment of LLMs' medical reasoning. MedThink-Bench comprises 500 challenging questions across ten medical domains, each annotated with expert-crafted step-by-step rationales. Building on this, we propose LLM-w-Ref, a novel evaluation framework that leverages fine-grained rationales and LLM-as-a-Judge mechanisms to assess intermediate reasoning with expert-level fidelity while maintaining scalability. Experiments show that LLM-w-Ref exhibits a strong positive correlation with expert judgments. Benchmarking twelve state-of-the-art LLMs, we find that smaller models (e.g., MedGemma-27B) can surpass larger proprietary counterparts (e.g., OpenAI-o3). Overall, MedThink-Bench offers a foundational tool for evaluating LLMs' medical reasoning, advancing their safe and responsible deployment in clinical practice.

Problem

• Evaluate medical reasoning of large language models rigorously
• Address scalability and fidelity gaps in current assessments
• Provide expert-level benchmark for clinical decision-making integration

Innovation

• Introduces MedThink-Bench for medical reasoning evaluation
• Proposes LLM-w-Ref with fine-grained rationale assessment
• Leverages LLM-as-a-Judge for scalable expert-level fidelity