Radiology Report Generation (RRG) has benefited from stronger encoder-decoder pipelines and particularly Large Language Model (LLM) report generators, but most methods still perform training through random mini-batches. This practice is poorly matched to chest radiology corpora, where common normal statements and frequent findings dominate the optimization signal, while clinically important pathological attributes appear in the long tail. We argue that random sampling pushes report generators toward a frequent-finding shortcut, where models quickly learn fluent report narratives and high-frequency disease phrases, then receive limited pressure to recover rare attributes that contribute little to aggregate cross-entropy. To address this problem, we propose ParaC, a pathology-aware replay curriculum for long-tail RRG. ParaC ranks training reports with pathology labels and structured medical knowledge, expands the training process from simple reports to complex long-tail cases, and replays prototype anchors to preserve basic cross-modal grounding. Extensive experiments on IU X-Ray and MIMIC-CXR confirm the superiority of ParaC on several existing methods, with further ablations analyzing the large gains brought by our pathology-aware curriculum to cover rare pathological findings.

Building fine-grained cross-modal alignment via pure radiograph-report pairs has been a long-standing challenge in the task of Radiology Report Generation (RRG). An intuitive solution towards this issue lies at introducing extra contexts to assist the report generation process, yet clinical contexts are normally unavailable in most real-world scenarios. In this paper, we present an asymmetric context self-distillation framework for RRG, namely ConDistill. To facilitate radiograph-only report generation, it refines the basic radiograph-report alignment, by progressively injecting additional contexts via multi-stage curriculum. Since these contexts are not available in inference, we develop a self-distillation strategy to transfer the bridge states from the teacher branches with extra contexts to the student one with solely radiograph inputs. For evaluation, we conduct extensive experiments on both radiograph-only and multi-context benchmarks, i.e., MIMIC-CXR and MIMIC-RG4, respectively. Experimental results validate our promising performance and further ablations offer potential insights for the context learning in LLM-based RRG frameworks.