LetTokenBurn
Sign in
Auto57%Harbor: pendingSafety & AlignmentAgentic / Tool Use

MediateRec

Personal-Agent Mediated Recommendation with Cross-Platform User History

A benchmark for personal-agent mediated recommendation using scalable proxy environments (Amazon Reviews) and a real cross-platform test to evaluate rescue-harm trade-offs.

Paper

How to run

ltb · mediaterec
$ npm i -g lettokenburn-cli
$ ltb bench info mediaterec

External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.

Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs

Scores

4 sourced entries
HarnessSourceConfidenceLink
Claude Opus 4.613.6N/AIndependentmediumOct 6, 2026arxiv.org
Claude Sonnet 4.613.5N/AIndependentmediumOct 6, 2026arxiv.org
Claude Haiku 4.512.2N/AIndependentmediumOct 6, 2026arxiv.org
Qwen3 4B8.5N/AIndependentmediumOct 6, 2026arxiv.org

Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.

Reproduced by the community

All runs
Loading community runs…

About

MediateRec constructs controlled platform–agent information boundaries using Amazon Reviews for proxy environments and linked activity data for real cross-platform testing. It evaluates how well personal agents use cross-platform history to correct platform rankings without introducing harmful overrides.

Scoring focuses on the balance between beneficial rescues and harmful overrides compared to platform baselines. Experiments show that while LLMs can make meaningful corrections, they still introduce non-negligible harmful overrides, highlighting the need for optimized mediation strategies like PAMO.