Med-R2: Perception and Reflection-driven Complex Reasoning for Medical Report Generation

A new fine-tuning strategy called Med-R2 aims to improve automated medical report generation by adding a perception-driven reasoning step before the final report is produced, according to a preprint posted on arXiv [1]. The approach targets a known weakness in current medical report generators built on large vision-language models (LVLMs). Most state-of-the-art systems rely on direct supervised fine-tuning with medical image-report pairs, a method that generates reports without an intermediate diagnostic reasoning process [1]. The authors argue this can cause models to miss pathological features and produce incorrect diagnoses [1]. Direct supervised fine-tuning also lacks radiology-specific knowledge guidance, which can lead to misinterpretation of perceived features [2]. Med-R2 introduces a perception-driven long reasoning process that precedes report generation and incorporates radiology-specific knowledge as guidance [2]. A reflection mechanism is added to refine both the perception of pathological features and the generated report, aiming to reduce perceptual errors during complex reasoning [2]. The preprint was submitted to arXiv on 2 April 2025 and revised on 18 June 2026 [1]. The initial submission was 1,248 KB, and the revised version is 740 KB [1]. arXiv, where the paper appears, is an open-access repository for electronic preprints that are moderated but not peer-reviewed [6]. Founded in 1991, the repository passed the two-million-article milestone by the end of 2021 and receives roughly 24,000 submissions per month as of late 2024 [6]. The paper is listed under the Computation and Language category and is accessible through arXiv's abstract page, which includes experimental community tools developed under the arXivLabs framework [1][5]. arXivLabs is a formalized collaboration space that allows third-party developers to build features on top of arXiv's article pages, such as bibliographic explorers and code recommenders [5]. The framework was launched in 2020 to encourage community innovation while ensuring partners adhere to arXiv's values of openness, community, excellence, and user data privacy [5]. The Med-R2 paper's abstract page displays several of these Labs integrations, including Bibliographic Explorer, Connected Papers, and CORE Recommender [1][4]. The authors report that experiments demonstrate Med-R2 effectively enhances pathological feature perception and diagnosis accuracy for medical report generation via fine-tuned LVLMs [2]. The work was authored by Hao Wang and colleagues [1].