Personal-Agent Mediated Recommendation with Cross-Platform User History

Personal-Agent Mediated Recommendation with Cross-Platform User History. By Yu Xia and four others, from the University of California San Diego and Meta AI. Published on the sixth of October. No code repository is linked to it.

Recommendation has long been centred on the platform: each service personalises using what it knows about you. The shift described here is toward personalisation governed by the user, where a personal LLM agent acts on the user's behalf across services. The paper formalises this as Personal-Agent Mediated Recommendation. A platform recommender ranks a set of candidates using only platform-local information. A personal agent then uses cross-platform history that the user has authorised, mediates that ranking, and produces the final top-K slate.

Mediation sounds simple and is not. The platform's ranking can encode strong evidence about the wider population, evidence the personal agent cannot see. So every change the agent makes is a bet. Some changes are beneficial rescues, surfacing something the platform missed. Others are harmful overrides, throwing away a ranking that was right. Effective mediation has to balance the two.

Two contributions answer this. MediateRec is a benchmark for studying the trade-off, and Personal Attribution Mediation Optimization, or PAMO, is a way of training the agent to use cross-platform history well. The idea behind PAMO is to ask how much a given mediation actually depends on the personal history, and to let that answer shape the learning signal.

MediateRec includes scalable proxy environments that span platforms, and a real cross-platform test, all under a controlled information boundary between platform and agent. PAMO counterfactually masks the cross-platform history to estimate how much personal support a mediation has. It then reallocates rank-aware advantage mass, subject to a value floor defined relative to the platform. The authors prove that PAMO preserves advantage mass at the cutoff, and that it is locally optimal among first-order reallocations that preserve this mass without lowering average value relative to the platform.

Experiments on MediateRec show that mediation by a personal agent does enable meaningful corrections to the platform. They also show the risk is real: even strong proprietary LLMs introduce harmful overrides that are not negligible. PAMO consistently improves over matched reinforcement learning that rewards outcomes only. That holds across seen and unseen target platforms and on the real cross-platform test, and PAMO achieves a better balance between rescues and harm.

The framing does much of the work. Once the personal agent is seen as editing a ranking built on evidence it cannot observe, the goal is no longer to change as much as possible but to change only where personal history genuinely supports it. PAMO turns that principle into a training signal, and MediateRec gives a place to measure whether it holds.