From Hidden Profiles to Governable Personalization: Recommender Systems in the Age of LLM Agents
New paper argues LLM agents like ChatGPT will fundamentally change how recommendation systems work.
A team of nine researchers led by Jiahao Liu has published a significant paper arguing that the rise of LLM agents (like ChatGPT, Claude, and Gemini) is forcing a fundamental redesign of recommendation systems. Traditionally, platforms like Netflix or Amazon build hidden, proprietary user profiles optimized for prediction but inaccessible to users. The paper contends that as AI assistants mediate more activities—from shopping to travel planning—user representation can no longer be confined to isolated platforms. This creates an opportunity and necessity to shift toward 'governable personalization,' where users can inspect, revise, and even port their digital profiles across services.
The researchers identify five interconnected research fronts essential for this new paradigm. First is creating transparent yet privacy-preserving user models. Second is 'intent translation and alignment,' ensuring AI assistants correctly interpret and act on user goals. Third involves designing cross-domain representation and memory systems that work across different platforms. Fourth addresses trustworthy commercialization in environments mediated by AI. Finally, the paper calls for new operational mechanisms for data ownership, access, and accountability. The team positions these not as isolated technical puzzles but as core design problems created by the emergence of LLM agents as universal intermediaries.
- Proposes shift from hidden, platform-specific user profiles to portable, user-inspectable representations controlled by individuals.
- Identifies five key research areas: transparent modeling, intent alignment, cross-domain memory, trustworthy commercialization, and governance mechanisms.
- Argues the problem is not just better AI inference, but building systems users can meaningfully understand, shape, and govern.
Why It Matters
This could fundamentally shift power from platforms to users, making AI recommendations more transparent and controllable.