Researchers: LLM RecSys must use explicit user feedback to beat filter bubbles

A new research paper from Zhang et al. (CogMI 2025) challenges the dominant approach in LLM-based recommendation systems (RecSys). Traditionally, these systems infer user preferences from implicit signals like clicks, purchases, or watch time. But the authors argue that this ignores the rich, explicit contextual feedback users provide through natural language—comments, reviews, and other written input. This 'explicit context feedback' captures the nuanced reasons behind user decisions, offering critical heterogeneous information for preference alignment. Overlooking such signals, they claim, leads to misaligned preferences and reinforces filter bubbles, as algorithms fail to understand the semantic context driving user choices.

The paper advocates for a paradigm shift: building the next generation of LLM-driven RecSys around explicit user signals rather than just item metadata or implicit behavior. The authors review how recommendation paradigms have evolved, highlight the underutilized value of user-generated text, and call for new benchmarks and metrics that reflect context-rich user input. They also introduce frameworks for integrating explicit feedback into scalable, LLM-based RecSys. Centering on user-preference modeling, the goal is to produce more personalized, transparent, and explainable recommendations that truly reflect why users like or dislike certain items—ultimately moving beyond the limitations of current systems.