An Analysis of Training-Free Self-Reported Confidence in Language Models

Summary

This workshop paper asks whether training-free confidence signals from black-box language models are more than calibrated rhetoric. On 99 audited TriviaQA questions, the authors compare three signals across two model families (DeepSeek Flash, Claude Sonnet 5): confidence verbalized with the answer, post-hoc P(True) for a fixed answer, and agreement among three additional samples. After manually auditing benchmark labels — correcting four DeepSeek and three Sonnet false negatives and excluding one disputed item — verbalized confidence reaches AUROC 0.956 and 0.937, P(True) 0.920 and 0.959, while three-sample agreement is much weaker (0.765, 0.790). A fixed 50/50 average of confidence and agreement shows no statistically reliable benefit (paired bootstrap intervals cross zero), and several audited errors receive unanimous sample support, showing self-consistency can amplify shared misconceptions. Re-eliciting confidence with semantically equivalent prompts for fixed answers shifts scores by 0.043–0.084 MAE and flips 4–9.1% of decisions at a 0.8 threshold. An exploratory audit of 100 biography claims finds mean confidence of 0.919 for supported versus 0.830 for contradicted claims, but only five claims are contradicted. Main limitations: only 8–9 errors per model, a single dataset and two model families, an agreement baseline limited to three samples with alias matching and no semantic clustering, uncontrolled Sonnet temperature, and unaudited supported claims.