MMedExpert-R1: Strengthening Multimodal Medical Reasoning via Domain-Specific Adaptation and Clinical Guideline Reinforcement

TL;DR

MMedExpert-R1 targets robust multimodal medical reasoning by addressing data, initialization, and reasoning diversity gaps in MedVLMs. It introduces a three-part framework: Domain-Specific Adaptation (DSA) with specialty LoRA adapters for diverse initialization, Guideline-Based Advantages (GBA) to align reasoning with clinical guidelines, and Conflict-Aware Capability Integration via a TIES-based merging strategy to unify domain experts with RL-based reasoning. The authors also present MMedExpert, a high-quality reasoning dataset with 3.9K annotated samples across AM, CAH, BS, and CSI, featuring step-by-step traces across four reasoning paradigms. Across four benchmarks and two model scales (2B and 7B), MMedExpert-R1 achieves state-of-the-art results, validating the approach and highlighting potential for reliable, guideline-consistent clinical decision support in multimodal settings.

Abstract

Medical Vision-Language Models (MedVLMs) excel at perception tasks but struggle with complex clinical reasoning required in real-world scenarios. While reinforcement learning (RL) has been explored to enhance reasoning capabilities, existing approaches face critical mismatches: the scarcity of deep reasoning data, cold-start limits multi-specialty alignment, and standard RL algorithms fail to model clinical reasoning diversity. We propose MMedExpert-R1, a novel reasoning MedVLM that addresses these challenges through domain-specific adaptation and clinical guideline reinforcement. We construct MMedExpert, a high-quality dataset of 10K samples across four specialties with step-by-step reasoning traces. Our Domain-Specific Adaptation (DSA) creates specialty-specific LoRA modules to provide diverse initialization, while Guideline-Based Advantages (GBA) explicitly models different clinical reasoning perspectives to align with real-world diagnostic strategies. Conflict-Aware Capability Integration then merges these specialized experts into a unified agent, ensuring robust multi-specialty alignment. Comprehensive experiments demonstrate state-of-the-art performance, with our 7B model achieving 27.50 on MedXpert-MM and 83.03 on OmniMedVQA, establishing a robust foundation for reliable multimodal medical reasoning systems.

Figures (5)

• Figure 1: The core elements in MMedExpert-R1. The framework combines Domain-Specific Adaptation (DSA) through specialty-aware LoRA modules, Clinical Reasoning Data from mixed medical cases across multiple domains, and Guideline-Based Advantage estimation guided by clinical reasoning guidelines.
• Figure 2: (a) Data construction of MMedExpert; (b) MMedExpert-R1 consists of a Domain-Specific Adaptation to provide diverse initialization, Guideline-Based Advantages to align with real-world diagnostic strategies, and a Conflict-Aware Capability Integration to ensure multi-specialty alignment.
• Figure 3: Ablation studies on MMedExpert-R1. (a) illustrates the impact of Domain-Specific Adaptation. (b) shows optimization of Conflict-Aware Capability Integration. (c) demonstrates the superiority of our merging technique.
• Figure 4: Visualization on MedXpert-MM.
• Figure 5: Visualization on MedXpert-MM.