RxnNano：通过分层课程学习训练紧凑型大语言模型以进行化学反应和逆合成预测

Chemical reaction prediction is crucial for accelerating drug discovery and synthesis planning. Despite advancements in data-driven models, current approaches are limited by an overemphasis on scaling parameters and datasets. Some methods, coupled with evaluation techniques, bypass fundamental challenges in reaction representation and fail to capture deep chemical intuitions such as reaction common sense and topological atom mapping logic. The core challenge lies in instilling this knowledge into the models. A unified framework is proposed that prioritizes chemical understanding over scale through three key innovations: (1) a Latent Chemical Consistency objective that models reactions as movements in a continuous chemical space; (2) a hierarchical curriculum learning strategy that gradually guides the model to learn complex chemical knowledge; and (3) a compact large language model design to enhance computational efficiency and prediction accuracy.