We prototype automatic evaluators that combine structured perturbations with learned reward models. The project explores stability of metrics under prompt variance and noisy annotations, and proposes lightweight calibration steps that significantly improve agreement with human judgments.
Research Project