IC-369LMs verbalizing confidence with certainty phrases are better calibrated on SCIQ than on TruthfulQA

Peiqi Wang, Barbara D. Lam, Yingcheng Liu, Ameneh Asgari-Targhi, Rameswar Panda, William M Wells, Tina Kapur, Polina Golland

SourceCalibrating Expressions of Certainty

All three large LMs evaluated show substantially lower ECE on SCIQ (1000 science exam questions) than on TruthfulQA (817 questions designed to test mimicry of human misconceptions). GPT-4o drops from ECE 0.07 to 0.22, Claude-3.5-Sonnet from 0.06 to 0.25, and Gemini-1.5-Pro from 0.17 to 0.28. Accuracy also drops sharply (e.g., GPT-4o from 0.73 to 0.35), suggesting the harder dataset exposes both lower discriminative ability and worse calibration.

Evidence
correlational
Key metric
GPT-4o: SCIQ ece 0.07 vs TruthfulQA ece 0.22; Claude-3.5-Sonnet: SCIQ ece 0.06 vs TruthfulQA ece 0.25; Gemini-1.5-Pro: SCIQ ece 0.17 vs TruthfulQA ece 0.28
Caveat
The calibration gap may be confounded by the large drop in accuracy on TruthfulQA (e.g., GPT-4o acc 0.73 vs 0.35), since calibration is harder to achieve when the model is less discriminative.
Model
GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 / Gemini Pro 1.5 Gemini 1.5 Pro
Datasets
SciQ / SciQA [eval], TruthfulQA / TruthfulQA MC1 [eval]
Related work
Tian et al. 2023 (Just Ask for Calibration) [context]
Related findings
IC-368
Extraction
automatic-extraction