IC-179GPT-4o mini, GPT-4o, and Llama-3-8B all calibrate confidence in their internal answers significantly better than confidence in external contexts

Yukun Huang, Sanxing Chen, Hongyi Cai, Bhuwan Dhingra

SourceTo Trust or Not to Trust? Enhancing Large Language Models' Situated Faithfulness to External Contexts

The paper measures expected calibration error (ECE) for the confidence scores each model assigns to its internal answer (internalconf) versus its context-based answer (contextconf) across six datasets. For Llama-3-8B, internal ECE ranges from 13.6 to 23.6 while context ECE ranges from 32.8 to 43.1. For GPT-4o mini, internal ECE is 7.7–25.9 versus context ECE 30.4–45.9. For GPT-4o, internal ECE is 6.5–27.6 versus context ECE 25.0–39.9. The gap is consistent across all three models and all datasets, indicating a systematic asymmetry in the models' ability to produce well-calibrated confidence for external information.

Evidence
correlational
Key metric
ECE: Llama-3-8B internal 13.6–23.6 vs context 32.8–43.1; GPT-4o mini internal 7.7–25.9 vs context 30.4–45.9; GPT-4o internal 6.5–27.6 vs context 25.0–39.9 (across TriviaQA, PopQA, NaturalQA, FreshQA, RedditQA, ClashEval)
Caveat
Calibration is measured on the confidence scores elicited by specific prompting procedures; the gap may narrow or widen under different elicitation methods. The paper notes that improving ECE via isotonic regression does not always translate to better situated faithfulness.
Model
GPT-4o mini, Llama 3 8B
Concepts
Failure mode
Datasets
TriviaQA [eval], PopQA [eval], Natural Questions / NaturalQA [eval], FreshQA [eval], RedditQA [eval], ClashEval [eval]
Related findings
IC-177, IC-178, IC-180
Extraction
automatic-extraction