Md Omar Faruk Rokon

Cost-efficient relevance evaluation can be achieved without sacrificing accuracy, thanks to a novel calibrated model cascade that halves compute costs while enhancing annotation quality.

Leveraging structured training signals from retrieval system disagreements, this approach boosts retrieval performance and ad engagement metrics significantly without traditional click-based methods.

The MHP Ensemble not only outperforms existing models but also reveals critical insights into how task routing architectures impact performance across different model types.