Pointwise, pairwise, or listwise - pick one for a promp...

A strong answer identifies the governing variable: how many documents the model sees per decision, which fixes call count, serial depth, and whether the model provides a calibrated scale or an ordering. A weak answer simply cites ‘state-of-the-art’ status for listwise, which fails to explain why pointwise dominates production rerankers—the latter are fine-tuned, allowing the coupling to live in the loss instead of the context.