Dynamic layer skipping reduces LLM computation by allowing each token to execute only a subset of the model's layers. However, existing skippers rely on spec…
机构:剑桥
来源:arXiv 2609.37062 | AI4Papers 论文推荐平台