benchmarksopinionbullishRegret-based evaluation metrics better reveal whether models clarify usefully and efficiently compared to simple accuracy measuresComputation and Language28 Jul 2026http://arxiv.org/abs/2607.21143v1