Personal-Agent Mediated Recommendation with Cross-Platform User History
The paper tests whether a user-controlled LLM agent can safely reorder platform recommendations using cross-platform history.
It defines “personal-agent mediated recommendation,” where the platform ranks candidates first and the agent adjusts the final top-K list under a controlled information boundary. The authors introduce MediateRec to measure that trade-off, including proxy cross-platform settings and a real cross-platform test. Their PAMO method uses counterfactual masking of user history to decide when personal evidence supports a change, aiming to preserve useful platform signal while reducing bad overrides. Experiments report better rescue-harm balance than matched outcome-only RL, though even strong proprietary LLMs still caused meaningful harmful overrides. Source: HF Daily Papers' note
It defines “personal-agent mediated recommendation,” where the platform ranks candidates first and the agent adjusts the final top-K list under a controlled information boundary. The authors introduce MediateRec to measure that trade-off, including proxy cross-platform settings and a real cross-platform test. Their PAMO method uses counterfactual masking of user history to decide when personal evidence supports a change, aiming to preserve useful platform signal while reducing bad overrides. Experiments report better rescue-harm balance than matched outcome-only RL, though even strong proprietary LLMs still caused meaningful harmful overrides. Source: HF Daily Papers' note
score 4