Recommender systems have long been designed for platforms, not for the people using them. According to a new arXiv abstract, this platform-first approach creates behaviors like clickbait and filter bubbles that serve platform interests but annoy users. The authors argue that such phenomena stem from a fundamental design choice.
Their proposed system, AgentRecommender, inverts that design. Instead of relying on the platform's algorithm, users deploy LLM agents that operate on the user's side. This allows for customizable recommender behavior that the user controls, rather than receiving whatever the platform decides to surface.
The abstract is brief, and the full paper is not yet available in the source. Still, the core idea is clear: recommender systems need not be a black box owned by the platform. By placing LLM agents with the user, the authors aim to give individuals the ability to shape their own recommendations and avoid the side effects of platform-centric ranking.