1 paper across 1 session
We propose LayerIF, a framework that employs Influence Functions for LLM layer quality estimation. Our method captures task-specific layer importance and improves both expert allocation in LoRA-MoE and layer-wise sparsity distribution in LLM pruning.