Improving LLM-Based Recommender Systems with User-Controllable Profiles

Large Language Models (LLMs) have demonstrated significant potential across various domains, including their application in recommendation systems (RS). In this paper, we propose a method that emphasizes user control, thereby increasing the role of the human within the system. Our research investigates the effectiveness of a variety of LLMs in capturing and using user preferences for recommendation tasks. The findings reveal that incorporating user controllability into RS can enhance performance by up to 50%. Furthermore, the results highlight that textual and user-controlled representations of preferences, called user-controllable profiles, outperform historical data to improve recommendation quality.