Meta formalizes personal-agent mediated recommendation, releases MediateRec benchmark
meta · hf · 2026-10-07
Meta's paper shifts recommendation from platform-centric to user-governed personalization: a personal LLM agent with user-authorized cross-platform history mediates the platform's ranking to produce the final top-K. The platform ranking encodes population evidence the agent can't see, so mediation must balance beneficial rescues against harmful overrides.
They release MediateRec (scalable proxy cross-platform environments plus a real cross-platform test under a controlled information boundary) and propose PAMO, which counterfactually masks cross-platform history to estimate personal mediation support and reallocates rank-aware advantage mass under a platform-relative value floor. Personal-agent mediation enables meaningful platform corrections, but even strong proprietary LLMs introduce non-negligible harmful overrides; PAMO beats outcome-only RL on seen and unseen platforms with a better rescue-harm balance.
More from Apps
- Cursor's iOS app now lets you monitor and direct coding agents running on your PC — XFreeze · 2026-10-07
- User lists five UX fixes for Dot agent: task visibility, unread noise, channels — carlosdponx · 2026-10-07
- Figure CEO Brett Adcock launches Hark Pro agent, demoed auto-applying for a Vietnam visa — adcock_brett · 2026-10-07
- Mamanoko's audio doubles as a basic Japanese learning module — TarlonKhoubyari · 2026-10-07
- Amazon's new AI feature that summarizes everything it knows about you goes viral — venturetwins · 2026-10-07
- Bops launches business-ops AI agents, each with its own computer, email and phone number — alexcovo_eth · 2026-10-07