KAIST AI 与 Google DeepMind 等发布论文《Language Models Can Control Their Own Attention》。