Precision Recall Controllable Radiology Report Generation via Hybrid Natural Language and Clinical Reward Learning

A research team has proposed a reinforcement learning framework that gives clinicians direct control over the precision-recall balance in automated radiology reports, aiming to align machine-generated text more closely with varying diagnostic needs [1]. Automated radiology report generation can ease the documentation burden on clinicians, but most existing systems optimize solely for language fluency metrics, leaving little room to adjust clinically important factors such as precision and recall [1][2]. The new framework, detailed in a paper revised on 23 June 2026, introduces a control parameter that explicitly adjusts the trade-off between clinical precision and recall during inference, allowing the model to flexibly generate reports according to different clinical requirements [1][2]. To improve clinical correctness, the authors incorporate a clinical reward into the training objective, moving beyond standard language-based optimization to boost clinical efficacy [1][2]. A group-relative training strategy normalizes rewards within each training batch, reducing reward variance and improving training stability [1][2]. Experiments on the MIMIC-CXR dataset showed the method consistently outperformed state-of-the-art approaches on both natural language generation and clinical efficacy metrics, while providing reliable control over the precision-recall trade-off [1][2]. The work arrives as large language models, typically based on transformer architectures, are increasingly applied to specialized domains such as medicine [3]. These models are pre-trained on vast text corpora and can generate, summarize, and analyze text, but biased or inaccurate training data can make outputs less reliable [3]. The new framework addresses that concern by tying generation objectives directly to clinical rewards rather than fluency alone [2]. The paper was posted on arXiv, the open-access repository that hosts electronic preprints across physics, computer science, and related fields [8]. arXiv, which began in 1991, surpassed two million articles by the end of 2021 and now receives about 24,000 submissions per month [8]. The platform also supports community-built tools through arXivLabs, a framework that enables collaborators to develop features such as citation explorers and code finders directly on article pages [6][7].