This paper addresses the challenge of radiology report generation (RRG) by proposing a method to bootstraps large language models (LLMs) for this task. The authors aim to improve the alignment between visual and textual features and generate more informative reports. They introduce two key components: in-domain instance induction and coarse-to-fine decoding. The in-domain instance induction process uses contrastive learning to align the LLM with radiology reports from general texts, while the coarse-to-fine decoding process refines intermediate reports to produce final, precise reports. Experimental results on two benchmark datasets, IU X-RAY and MIMIC-CXR, demonstrate the effectiveness of their approach, outperforming existing state-of-the-art methods. The study highlights the importance of domain adaptation and task-specific generation in RRG, providing a practical framework for future research in this area.This paper addresses the challenge of radiology report generation (RRG) by proposing a method to bootstraps large language models (LLMs) for this task. The authors aim to improve the alignment between visual and textual features and generate more informative reports. They introduce two key components: in-domain instance induction and coarse-to-fine decoding. The in-domain instance induction process uses contrastive learning to align the LLM with radiology reports from general texts, while the coarse-to-fine decoding process refines intermediate reports to produce final, precise reports. Experimental results on two benchmark datasets, IU X-RAY and MIMIC-CXR, demonstrate the effectiveness of their approach, outperforming existing state-of-the-art methods. The study highlights the importance of domain adaptation and task-specific generation in RRG, providing a practical framework for future research in this area.