IC-1265Calibration and failure prediction improve as model capability scales from GPT-3 to GPT-4, but remain far from ideal

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

Comparing vanilla verbalized confidence across GPT-3 (175B), GPT-3.5, and GPT-4 reveals a consistent trend: as model capability increases, ECE decreases and AUROC increases. The paper reports an approximate 22.2% improvement in AUROC from GPT-3 (avg 51.3) to GPT-4 (avg 62.7). However, even GPT-4's AUROC remains close to the 50% random-guess threshold, and its ECE of 18.0 (×10²) still indicates substantial miscalibration. The improvement is monotonic across the GPT family but does not reach satisfactory performance.

Evidence
correlational
Key metric
Approximate 22.2% improvement in AUROC from GPT-3 to GPT-4; average ECE (×10²): GPT-3 52.0, GPT-3.5 37.7, GPT-4 18.0; average AUROC: GPT-3 51.3, GPT-3.5 55.1, GPT-4 62.7
Caveat
The comparison spans different model families (GPT-3/3.5/4, Vicuna, Llama 2) rather than a single family's size variants, so the scaling trend conflates architecture changes with scale. The paper itself notes results remain 'far from ideal performance.'
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
Scale-dependent behaviour
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 findings
IC-1264, IC-1266
Extraction
automatic-extraction