With the rapid development of wide−field surveys, high−cadence monitoring, and multimessenger observations, astronomy is facing a new era of simultaneous growth in data volume, data variety, and observing frequency. At the same time, artificial intelligence is expanding from task−specific models to foundation models and large−language−model−based scientific agents. This article reviews representative advances of these methods in astrophysics over the past five years, with a focus on comparing their functional boundaries in terms of training objectives, data organization, transfer strategies, and scientific roles. The review shows that task−specific models remain the most mature and effective for well−defined problems with stable data distributions; foundation models, pretrained on large−scale heterogeneous data, have demonstrated promising potential for unified representation and cross−task transfer in imaging, spectroscopy, and time−series tasks; scientific agents, however, are still in early exploratory stages, and their roles in literature summarization, tool invocation, and multi−step research workflow orchestration require cautious assessment. The article further examines key challenges, including data quality, generalization, physical interpretability, uncertainty quantification, and reproducibility. It identifies future priorities in building high−quality data infrastructure, establishing cross−survey out−of−distribution validation standards, solidifying software toolchains, and fostering interdisciplinary talent training.