Beyond Adjacent Layers: Graph-Guided Layer Fusion for Compressing Large Language Models
FCM identifies non-adjacent functional communities within large language models by using full pairwise Centered Kernel Alignment similarity matrices. It groups layers based on functional equivalence, selects the most important layer from each group using Block Influence, and compresses the model by retaining only these representative layers. This method achieves superior compression gains and improves downstream performance compared to existing techniques.