IC-327Llama-3-8B-Instruct, Mistral-7B-Instruct-v0.3, and several other LLMs produce well-calibrated verbal confidence estimates on classification tasks

Brian Formento, Chuan-Sheng Foo, See-Kiong Ng

SourceConfidence Elicitation: A New Attack Vector for Large Language Models

The paper measures how well each model's self-reported verbal confidence (highest/high/medium/low/lowest) aligns with actual correctness using ECE and AUROC, averaged over 500 examples. Llama-3-8B-Instruct achieves ECE of 0.1264 (SST-2), 0.1376 (AG-News), and 0.0492 (StrategyQA) with AUROC above 0.66 on all three. Mistral-7B-Instruct-v0.3 shows similar calibration. Appendix F.1 extends this to Gemma2, Mistral-Nemo, Qwen2.5, and Llama-3.2-11B, all showing low ECE. GPT-4o shows the lowest ECE (0.0286 on SST-2). The authors conclude that well-calibrated models for confidence elicitation already exist and this capability is likely to improve with scale.

Evidence
correlational
Key metric
ECE: Llama-3-8B-Instruct 0.1264/0.1376/0.0492 (SST-2/AG-News/StrategyQA), AUROC 0.9696/0.9293/0.6607; Mistral-7B-Instruct-v0.3 0.1542/0.1216/0.1295, AUROC 0.9537/0.8826/0.6358; GPT-4o 0.0286/0.0641/0.2300, AUROC 0.9713/0.9306/0.7410
Caveat
Calibration is measured in a zero-shot classification setting with a specific two-step prompting strategy (k guesses then verbal confidence). The StrategyQA AUROC is notably lower (0.66 for Llama-3) than SST-2 (0.97), suggesting calibration degrades on harder reasoning tasks.
Model
Llama 3 8B Instruct, Mistral 7B / Mistral / Mistral 3 7B / Mistral-0.2-7B / Mistral-v0.1 Mistral-7B-Instruct-v0.3, GPT-4o, Gemma 2 Gemma-2 9B Instruct, Mistral-Nemo 12B-Instruct-2407, Qwen2.5 Qwen2.5-7B-Instruct, Llama-3-2-Vision Llama-3.2-11B Vision-Instruct
Datasets
SST-2 [eval], AG-News [eval], StrategyQA [eval], RTE [eval], QNLI [eval]
Methods
Expected Calibration Error / Integral Calibration Error (ECE) [eval], AUROC [eval]
Related work
Xiong et al. 2024 (Can LLMs Express Their Uncertainty?) [builds-on], Tian et al. 2023 (Just Ask for Calibration) [builds-on]
Related findings
IC-328, IC-329
Extraction
automatic-extraction