Can We Integrate Items into Models? Knowledge Editing to Align LLMs with Product Catalogs

Abstract

This study explores the potential of knowledge editing techniques to enhance Large Language Models (LLMs) for Conversational Recommender Systems (CRS). While LLMs like GPT, Llama, and Gemini have advanced conversational capabilities, they face challenges in representing dynamic, real-world item catalogs, often leading to inaccuracies and hallucinations in recommendations. This research preliminarily investigates whether knowledge editing can address these limitations by updating the internal knowledge of LLMs, thereby improving the accuracy of product information without full model retraining. Using the open-source Llama2 model, we apply two knowledge editing methods (GRACE and r-ROME) on a dataset of notebook listings. Our findings demonstrate improvements in the model’s ability to accurately represent product features, with r-ROME achieving the highest gain, while not decreasing model efficiency. The study highlights the perspective of utilizing knowledge editing to enhance CRS and suggests future work to explore broader applications and impacts on recommender systems performance.