Educating fashions to specific their uncertainty in phrases

We present {that a} GPT-3 mannequin can study to specific uncertainty about its personal solutions in pure language—with out use of mannequin logits. When given a query, the mannequin generates each a solution and a stage of confidence (e.g. “90% confidence” or “high confidence”). These ranges map to possibilities which can be nicely calibrated. The mannequin additionally stays reasonably calibrated below distribution shift, and is delicate to uncertainty in its personal solutions, moderately than imitating human examples. To our data, that is the primary time a mannequin has been proven to specific calibrated uncertainty about its personal solutions in pure language. For testing calibration, we introduce the CalibratedMath suite of duties. We examine the calibration of uncertainty expressed in phrases (“verbalized probability”) to uncertainty extracted from mannequin logits. Each sorts of uncertainty are able to generalizing calibration below distribution shift. We additionally present proof that GPT-3’s skill to generalize calibration is determined by pre-trained latent representations that correlate with epistemic uncertainty over its solutions.