Modern recommender systems are moving from platform‑centric personalization to user‑governed personalization, where a personal LLM agent acts on behalf of the user across services. We formalize this emerging paradigm as Personal‑Agent Mediated Recommendation: a platform recommender ranks candidates using only platform‑local signals, while a personal agent leverages user‑authorized cross‑platform history to mediate the ranking and produce the final top‑K slate.\ \ Mediation is not a trivial overwrite. Platform rankings often encode strong population evidence that the personal agent cannot observe; thus effective mediation must balance beneficial rescues against harmful overrides.\ \ To study this trade‑off we introduce the MediateRec benchmark, which provides scalable proxy cross‑platform environments and a real cross‑platform test under a controlled platform‑agent information boundary.\ \ For training the agent to use cross‑platform history effectively we propose Personal Attribution Mediation Optimization (PAMO). The key idea is to counterfactually mask history to estimate personal mediation support, then reallocate rank‑aware advantage mass while respecting a platform‑relative value floor. Formally, let $v_i$ denote the platform’s relative value for the $i$‑th item and @@@MATH_BLOCK2@@@ the advantage mass allocated by mediation; PAMO enforces: $$ \sum{i=1}^K \Delta_i = \text{advantage mass}, \quad \forall i, \; v_i + \Delta_i \geq \tau $$ where $\tau$ is the platform value floor. We prove that PAMO preserves cutoff‑level advantage mass and is locally optimal among all first‑order reallocations that keep the average platform‑relative value from decreasing.\ \ Experiments on MediateRec show that personal‑agent mediation can meaningfully correct platform recommendations, yet even strong proprietary LLMs introduce non‑negligible harmful overrides. Compared with outcome‑only RL, PAMO consistently improves performance on both seen and unseen target platforms and on the real cross‑platform test, while achieving a better rescue‑harm balance.
Review