Continuous Autoregressive Language Models

Chenze Shao, Darren Li, Fandong Meng +1

The efficiency of large language models (LLMs) is fundamentally limited by their sequential, token-by-token generation process. We argue that overcoming this bottleneck requires a…