More Than a Score: Reinforcement Learning with Verbalized Feedback for Generative Recommendation
Zipeng Chen, Xinyu Lin, Yuliang Sun, Sheng Zhang, Zeyu Wang, Xiaowen Shi, Junwei Pan, Chengguo Yin
原文摘要
Large language models (LLMs) are increasingly used for generative recommendation, where recommended items are generated directly from users' interaction histories. To further improve recommendation quality, reinforcement learning (RL) has been applied to align these models with recommendation objectives. However, existing RL methods rely on scalar rewards to implicitly guide exploration. These rewards do not specify which aspects of a candidate conflict with the user's preferences or how the next candidate should differ. To address this gap, we propose VerRec, a verbalized feedback-guided RL framework for generative recommendation. Specifically, a user simulator generates personalized verbalized feedback based on the user's historical preferences and the candidate item profile. The policy uses this feedback to produce refined recommendations. Original and refined candidates form an expanded group and are jointly optimized with group relative policy optimization. The user simulator is only required during training and is removed at inference time. Experiments on three Amazon review datasets with two generative recommendation backbones demonstrate that VerRec improves recommendation accuracy and ranking quality. Further analyses show that verbalized guidance encourages exploration of more diverse candidates and helps the policy discover more rewarding recommendations during training.
cs.IR