Ruochen Mao, Yuling Shi, Xiaodong Gu +1
Aligning large language models with human preferences is critical for creating reliable and controllable AI systems. A human preference can be visualized as a high-dimensional vect…