Sycophancy Detection on Sycophancy benchmark (full evaluation set)
0.732AUROCHypocrisy Gap
Evaluation results
| Method | Links | |
|---|---|---|
| Hypocrisy GapPredictor=H, Base Model=Gemma-2B-IT2026.01 | 0.732 | |
| Hypocrisy GapPredictor=H, Model=Gemma-2B-IT2026.01 | 0.731 | |
| Hypocrisy GapPredictor=H, Base Model=Llama-3.1-8B-Instruct2026.01 | 0.588 | |
| Hypocrisy GapPredictor=H, Model=Llama-3.1-8B-Instruct2026.01 | 0.587 | |
| Hypocrisy GapPredictor=H, Model=Qwen3-1.7B2026.01 | 0.549 | |
| Hypocrisy GapPredictor=H, Base Model=Qwen3-1.7B2026.01 | 0.549 | |
| Log-probability baselinePredictor=ΔLP, Base Model=Llama-3.1-8B-Instruct2026.01 | 0.5 | |
| log-probability baselinePredictor=Baseline, Model=Llama-3.1-8B-Instruct2026.01 | 0.499 | |
| log-probability baselinePredictor=Baseline, Model=Qwen3-1.7B2026.01 | 0.453 | |
| Log-probability baselinePredictor=ΔLP, Base Model=Qwen3-1.7B2026.01 | 0.452 | |
| Log-probability baselinePredictor=ΔLP, Base Model=Gemma-2B-IT2026.01 | 0.424 | |
| log-probability baselinePredictor=Baseline, Model=Gemma-2B-IT2026.01 | 0.421 |