IC-557Guard models show significantly degraded calibration under jailbreak attacks, with prompt classification ECE substantially higher than response classification ECE

Hongfu Liu, Hengguan Huang, Xiangming Gu, Hao Wang, Ye Wang

SourceOn Calibration of LLM-based Guard Models for Reliable Content Moderation

On the HarmBench-adv set containing adversarial prompts from jailbreak attacks (GCG, AutoDAN), all guard models exhibit higher ECE for prompt classification than for response classification. WildGuard achieves the best F1 score of 92.8% on prompt classification but still has an ECE of 34.9%, indicating that high accuracy does not imply reliable confidence. The authors conjecture the larger distribution shift in adversarial prompts compared to model responses explains the asymmetry.

Evidence
correlational
Key metric
WildGuard prompt classification: F1 92.8%, ECE 34.9% on HarmBench-adv; prompt classification ECE generally higher than response classification ECE across all 6 prompt-classifying models
Caveat
HarmBench-adv is a validation set for fine-tuning Llama2-variant classifiers in HarmBench; the adversarial attacks tested are GCG and AutoDAN specifically
Model
Llama Guard, Llama Guard 2, Llama-Guard 3, Aegis-Guard-Defensive, WildGuard, Mistral 7B / Mistral / Mistral 3 7B / Mistral-0.2-7B / Mistral-v0.1 HarmBench-Mistral
Concepts
Failure mode
Datasets
HarmBench / HarmBench Prompt / HarmBench Response / HarmBench-adv [eval]
Methods
Expected Calibration Error / Integral Calibration Error (ECE) [eval]
Related work
HarmBench / HarmBench Prompt / HarmBench Response / HarmBench-adv [context]
Related findings
IC-556, IC-558, IC-559
Extraction
automatic-extraction