AdaDSF Framework for Similarity-Driven LLM Inference Acceleration

AdaDSF converts pre-trained LLMs into depth-sparse models by allocating resources based on cosine similarity between layer inputs and outputs. The framework uses a lightweight router to skip redundant layers and tokens without requiring full retraining.