Go to AAMAS 2026 Conference homepageEnabling User Agency in Scalable Content Recommendations with Large Language Models
Keywords: Personal Agents, Interpretable Recommendations, Recommender Systems, Large Language Models
TL;DR: Proposes a user-centric recommendation system where users maintain their own interest profiles using their agents locally instead of content providers storing user data centrally.
Abstract: Traditional content recommender systems depend on centrally stored interaction histories, such as impressions and clicks, to train their models. This raises privacy concerns and disadvantages newer content providers who lack sufficient user data. We propose a user-centric alternative where personal agents build interpretable and editable profiles composed of natural language profile items. Each profile item has a learnable weight indicating its importance, and profiles are learned offline under the user’s control, so no behavioral data needs to be shared with content providers. Recommendations are generated by matching content with weighted profile item embeddings in a shared embedding space. This space is constructed using only content data: potential user interests are extracted from each item with large language models, and the space is fine-tuned for efficient retrieval. The approach shifts computation and data control to the user while maintaining the efficiency of modern recommender systems, since online recommendation reduces to approximate nearest-neighbor search in the shared embedding space. It also lowers barriers for new providers and mitigates cold-start issues. Experiments on MIND and Goodreads datasets demonstrate that our system outperforms established baselines while providing transparency, editability, and full user control over their profiles and data, reimagining personalized recommendation as a process owned by the user.
Area: Innovative Applications (IA)
Generative A I: I acknowledge that I have read and will follow this policy.
Submission Number: 1237
Loading