Bayesian Assessment of Sycophancy on BASIL User belief setting 1.0 (test)
0.156Bayesian Error (RMSE)gpt-4o-mini
Evaluation Results
| Method | Links | |
|---|---|---|
| gpt-4o-miniConfidence Elicitation Method=Direct Probing, Number of Samples=52025.08 | 0.156 | |
| gpt-4o-miniConfidence Elicitation Method=Direct Probing, Number of Samples=12025.08 | 0.184 | |
| llama-3.2:1bConfidence Elicitation Method=Hybrid2025.08 | 0.219 | |
| claude-haiku-4-5Confidence Elicitation Method=Direct Probing, Number of Samples=52025.08 | 0.244 | |
| phi-4:14bConfidence Elicitation Method=Direct Probing, Number of Samples=52025.08 | 0.246 | |
| gpt-4o-miniConfidence Elicitation Method=Hybrid2025.08 | 0.258 | |
| phi-4:14bConfidence Elicitation Method=Direct Probing, Number of Samples=12025.08 | 0.273 | |
| claude-haiku-4-5Confidence Elicitation Method=Direct Probing, Number of Samples=12025.08 | 0.273 | |
| llama-3.2:1bConfidence Elicitation Method=Direct Probing, Number of Samples=52025.08 | 0.316 | |
| llama-3.2:3bConfidence Elicitation Method=Direct Probing, Number of Samples=52025.08 | 0.32 | |
| llama-3.2:3bConfidence Elicitation Method=Direct Probing, Number of Samples=12025.08 | 0.33 | |
| llama-3.2:3bConfidence Elicitation Method=Hybrid2025.08 | 0.339 | |
| mistral:7bConfidence Elicitation Method=Direct Probing, Number of Samples=52025.08 | 0.341 | |
| llama-3.2:1bConfidence Elicitation Method=Direct Probing, Number of Samples=12025.08 | 0.366 | |
| mistral:7bConfidence Elicitation Method=Direct Probing, Number of Samples=12025.08 | 0.422 | |
| phi-4:14bConfidence Elicitation Method=Hybrid2025.08 | 0.425 | |
| claude-haiku-4-5Confidence Elicitation Method=Hybrid2025.08 | 0.476 | |
| mistral:7bConfidence Elicitation Method=Hybrid2025.08 | 0.477 |