IC-1264LLMs are overconfident when verbalizing confidence, with values concentrated in 80–100% and multiples of 5, yielding high ECE across all five tested models

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

When prompted to state their confidence in an answer, all five LLMs (GPT-3, GPT-3.5, GPT-4, Vicuna 13B, Llama 2 70B) produce verbalized confidence values that cluster in the 80%–100% range and are almost exclusively multiples of 5, mimicking human confidence expressions. The resulting expected calibration error is high: average ECE ranges from 52.0 (GPT-3) to 18.0 (GPT-4) across eight datasets. Even GPT-4, the best-calibrated model, shows an average AUROC of only 62.7%, close to the 50% random-guess threshold, indicating poor ability to distinguish correct from incorrect predictions.

Evidence
observational
Key metric
Average ECE (×10²): GPT-3 52.0, Vicuna 46.1, Llama 2 43.6, GPT-3.5 37.7, GPT-4 18.0; GPT-4 average AUROC 62.7; confidence values predominantly in 80%–100% range, multiples of 5
Caveat
The paper notes that the overconfidence pattern may reflect imitation of human confidence expressions in the training corpus (citing zhou et al., 2023), and that the evaluation is limited to fixed-form and free-form QA tasks with unique ground-truth answers.
Model
GPT-3 / GPT base, GPT-3.5 / ChatGPT-3.5, GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report, Vicuna, Llama 2 / Llama 2 base
Concepts
Failure mode
Datasets
GSM8K [eval], SVAMP [eval], StrategyQA [eval], MMLU / MMLU-Math [eval]
Methods
Verbalized Confidence / Verbal confidence elicitation [primary], Expected Calibration Error / Integral Calibration Error (ECE) [eval], AUROC [eval]
Related work
Zhou et al. 2023 (Navigating the Grey Area) [context], Lin et al. 2022 (Teaching Models to Express Their Uncertainty in Words) [context]
Related findings
IC-1265, IC-1266
Extraction
automatic-extraction