User-Controlled Intent Layers for LLM-Mediated Personalization: A Research Agenda for Recommender Systems

Abstract

Recommender systems deployed at scale are predominantly organized around platform-centric architectures in which each service independently constructs and optimizes an internal representation of the user. As large language models (LLMs) become upstream entry points to digital services, personalization may be reorganized around a different architecture: an assistant can mediate user intent and coordinate recommendation and action across multiple downstream platforms. We argue that the central transformation is a redistribution of representational control. Building on prior work on general, scrutable, and user-controllable models, we define a user-controlled intent layer as a user-addressable representation and governance layer between a user/assistant and downstream services, through which personalization intents can be inspected, edited, scoped, authorized, operationalized, revoked, and audited. It is not a single interface, memory module, profile schema, or protocol, although implementations may combine all of these components. We distinguish this layer from both platform profiles and LLM profiles, illustrate its lifecycle in an end-to-end recommendation scenario, and develop a research agenda across five dimensions: privacy-preserving profile transparency, intent alignment and platform negotiation, cross-domain representation and memory, trustworthy commercial architectures, and operational profile governance. We further identify latency, computational cost, and user effort as cross-cutting feasibility constraints. Our central claim is that agent-mediated personalization will depend not only on better prediction, but also on whether users can meaningfully govern how their intent is represented and acted upon across services.