Activity Cliffs are characterized as phenomenon wherein subtle molecular variations result in markedly different binding affinities to the same biological target. Existing single-dimensional and coarse-grained molecular-target modeling methods are unable to capture these subtle differences, thus failing to understand the significant changes caused by them. To address the above problem, we present a novel Cross-Dimensional Representation Optimization Self-Supervised Learning Framework (CROSS). Through multiple rounds of self-supervised learning, CROSS can integrate fine-grained feature information from atoms and motifs, and apply cross-dimensional information fusion to molecular representations. We posit that iteratively optimizing molecular representations during the upstream representation learning process contributes to acquiring more discriminative and knowledgeable molecular representation vectors. More specifically, we design an innovative individual molecular discrimination task as a pretext task. Then, we proposed a contrastive learning training architecture called LIM-MOCO, which eliminates the need for explicit data augmentation in molecules. Experimental evaluation on benchmark datasets shows that our strategies are both efficient and advantageous.