Before an agent acts, it must decide whether action is warranted. A single accuracy score, the authors of a new arXiv paper argue, cannot tell whether a model failed to recognize that action was needed or recognized it but was reluctant to act.
Signal-detection theory separates these two aspects: sensitivity is the ability to distinguish warranted from unwarranted situations, while criterion is the threshold for acting. The paper, titled "Sensitivity or Criterion? Measuring and Moving When Language Models Act," applies this framework to language models.
By separating the two, the framework could help researchers identify whether a model's poor performance stems from perception or decision policy, and then adjust the appropriate component. The abstract does not detail the proposed measurement or intervention methods, so the full contribution remains to be seen.