Abstract
Content recommendation systems rely on user-behavior statistics to model preferences. These systems must be engineered to handle an ever-growing corpus of documents, a vast user base, and sparse signals for inferring latent interests. Additionally, they face significant challenges in determining the appropriate dimensionality and granularity for interest modeling. To produce high-quality systems that function at web-scale, it is essential to leverage global semantic knowledge. Conventionally, this is achieved through various embedding and tagging systems that provide a foundational vocabulary for interest modeling. In this work, we propose a multi-stage, query-based architecture that leverages Large Language Models (LLMs) to learn high-quality interest representations, explore novel interests, and optimize content presentation. These queries form a highly flexible vocabulary naturally suited for LLMs, while simultaneously facilitating human interpretability and tunability. Finally, we present protocols for prompt-based and gradient-based adaptation to optimize precision and recall over available user interaction data.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Sinha, Animesh; Patel, Jital; and Gaur, Akshay, "Tunable Multi-Stage Query-Based Protocol for Personalized Content Recommendations", Technical Disclosure Commons, (February 12, 2026)
https://www.tdcommons.org/dpubs_series/9325