Uncertainty and calibration

Steering a model's internal confidence moves how often it refuses to answer

Researchers publishing in Nature intervened directly on the confidence representation inside language models and watched abstention move the other way, making a refusal rate something with a control input, at least on multiple-choice facts.