Confidence Estimation on HH-RLHF
0.4718Rank Correlation (RK)Predictive Probability
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Predictive ProbabilityModel=Llama3-7B2026.05 | 0.4718 | 0.5294 | |
| Verbalized ConfidenceModel=Llama3-7B2026.05 | 0.4621 | 0.5336 | |
| Predictive ProbabilityModel=Qwen2.5-32B2026.05 | 0.4524 | 0.5483 | |
| Simulated AnnotatorsModel=Llama3-7B2026.05 | 0.391 | 0.6047 | |
| Random AnnotatorModel=Llama3-7B2026.05 | 0.3854 | 0.6199 | |
| Random AnnotatorModel=Qwen2.5-32B2026.05 | 0.3806 | 0.6279 | |
| Learning Confidence (Vanilla)Model=Llama3-7B2026.05 | 0.3766 | 0.624 | |
| Learning Confidence (Vanilla)Model=Qwen2.5-32B2026.05 | 0.3658 | 0.6392 | |
| Simulated AnnotatorsModel=Qwen2.5-32B2026.05 | 0.3629 | 0.6485 | |
| Learning Confidence (Ours)Model=Llama3-7B2026.05 | 0.3305 | 0.6784 | |
| Learning Confidence (Ours)Model=Qwen2.5-32B2026.05 | 0.3123 | 0.6912 |