IC-1266For GPT-3, white-box token-probability methods outperform black-box verbalized confidence in uncertainty estimation, but the gap is narrow (0.522–0.605 AUROC) and both remain near random

Miao Xiong, Zhiyuan Hu, Xinyang Lu, YIFEI LI, Jie Fu, Junxian He, Bryan Hooi

SourceCan LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMs

On GPT-3, three token-probability-based white-box methods (sequence probability, length-normalized sequence probability, key token probability) are compared against black-box verbalized confidence across five datasets. White-box methods consistently achieve lower ECE and higher AUROC, with length-normalized sequence probability and key token probability being most effective. However, the AUROC gap between the best white-box and best black-box method is only 0.522 to 0.605, and nearly all methods across datasets range between 0.5 and 0.6 in AUROC, indicating that neither access mode yields reliable uncertainty estimates from GPT-3.

Evidence
correlational
Key metric
AUROC gap between white-box and black-box: 0.522 to 0.605; nearly all methods across datasets range between 0.5–0.6 in AUROC
Caveat
The comparison is limited to GPT-3 (the only model with accessible token probabilities in the study) and five datasets. The paper notes that logit-based methods 'predominantly capture the model's uncertainty regarding the next token' rather than semantic uncertainty.
Model
GPT-3 / GPT base
Datasets
GSM8K [eval], StrategyQA [eval], MMLU / MMLU-Math [eval], SVAMP [eval]
Methods
Verbalized Confidence / Verbal confidence elicitation [compared-to], AUROC [eval], Expected Calibration Error / Integral Calibration Error (ECE) [eval]
Related work
Kuhn et al. 2023 (Semantic Uncertainty) [context]
Related findings
IC-1264, IC-1265
Extraction
automatic-extraction