“…Astonished by the miracle exhibited by LLMs, there has sparked a trend in contemplating how to utilize the power of LLMs in the field of recommendation system [8,23,28,42,46,52,59]. Current research on LRS follows a general pipeline: translating recommendation data into natural language input and then utilizing LLMs to generate recommendation results in a natural language form [15,26,29,56,57]. However, due to limitations such as the lack of recommendation data during the pre-training phase of LLMs, directly using LLMs for recommendation can only achieve suboptimal performance, making it necessary to tune LLMs on the recommendation data to unleash their recommendation capabilities [1,2,54].…”