Retrieval-augmented and feedback-optimized large language model for recommendation.

Abstract

Large Language Models (LLMs), with their powerful generative capabilities, have given rise to the emerging paradigm of "LLM-as-Recommender Systems" (LLM-as-RS). However, challenges such as data noise, the static nature of LLMs, and input length limitations hinder their full potential and make model updates costly. To address these issues, this study proposes an innovative generative recommendation framework: KDRAG-Critic-LLM-RS, which integrates retrieval augmentation and feedback optimization. Specifically, we design an efficient KDTreeRAG retrieval module, based on the principle of user-based collaborative filtering, to rapidly identify similar users and provide structured behavioral context. In addition, we introduce a lightweight R-Critic feedback module, inspired by item-based collaborative filtering, which captures user feedback signals to refine the generated recommendations and improve adaptability. By establishing a closed-loop mechanism of retrieval-generation-feedback, this framework effectively integrates collaborative signals from both user and item perspectives, continuously optimizing recommendation quality. Extensive experiments on real-world movie and book datasets demonstrate that KDRAG-Critic-LLM-RS significantly outperforms traditional baselines in terms of accuracy, personalization, and diversity-without requiring any fine-tuning of the LLM. More importantly, the framework is modular and plug-and-play, allowing the retrieval and feedback components to be easily integrated into any LLM-based recommendation system. This design avoids the high cost of frequent LLM retraining, mitigates the impact of noisy data, and overcomes input length limitations, ultimately enhancing the system's responsiveness to dynamic user preferences and improving generalization.