cross-layer-injection

The article introduces Echo-LoRA, a new parameter-efficient fine-tuning method that injects cross-layer representations from deeper source layers into shallow LoRA modules to improve performance without adding inference-time overhead.