Abstract:Large pre-trained models have achieved excellent performance in tasks such as natural language processing and computer vision, but still require fine-tuning when adapting to specific downstream tasks. As model sizes continue to grow, existing parameter-efficient fine-tuning (PEFT) methods still incur substantial computational and storage costs during both training and inference, limiting their applicability in resource-constrained environments. To address this problem, this study proposes a cross-layer PEFT method based on grouped parameter sharing, which further compresses the number of trainable parameters while maintaining model performance, thus achieving lightweight fine-tuning. First, the role of parameter sharing in PEFT is systematically analyzed, revealing the underlying reasons for balancing parameter compression and maintaining model performance through cross-layer parameter sharing. Second, the centered kernel alignment (CKA) metric is introduced to quantitatively analyze the similarity of hierarchical representations, and a grouping strategy is designed based on inter-layer correlations. Finally, a cross-layer parameter sharing mechanism is designed, where a set of parameters is shared within functionally similar layer groups, and learnable scaling factors are introduced to characterize hierarchical differences, enabling lightweight fine-tuning while maintaining adaptation performance. Experimental results show that on 19 downstream tasks of the VTAB benchmark, the proposed method substantially reduces the number of parameters (e.g., LoRA parameters can be reduced to one-third of the original), achieves the best performance on 14 tasks, and performs comparably to existing PEFT methods on the remaining tasks, further verifying the effectiveness and practicality of the proposed method.