Downstream utility of the genuine-followup metric

Establish the practical utility of the genuine-followup metric for downstream applications, including best-of-N assistant response selection, response reranking, and the construction of self-play training datasets.

Background

The work demonstrates that the genuine-followup metric captures a dimension of interaction awareness distinct from task accuracy and that it responds to perturbations and post-training interventions.

However, the authors have not yet demonstrated how this metric can be operationalized for downstream tasks such as selecting among multiple assistant responses, reranking outputs, or generating synthetic training data via self-play. They explicitly defer this to future work.