Hypocritical Sycophancy Detection on Sycophancy dataset knows-truth
74AUROCHypocrisy Gap (H)
Evaluation Results
| Method | Links | |
|---|---|---|
| Hypocrisy Gap (H)backbone=Gemma-2B-IT2026.01 | 74 | |
| Hypocrisy GapPredictor=H, Model=Gemma-2B-IT2026.01 | 73.9 | |
| Hypocrisy Gap (H)backbone=Llama-3.1-8B-Instruct2026.01 | 55.9 | |
| Hypocrisy GapPredictor=H, Model=Llama-3.1-8B-Instruct2026.01 | 55.8 | |
| Hypocrisy GapPredictor=H, Model=Qwen3-1.7B2026.01 | 55 | |
| Hypocrisy Gap (H)backbone=Qwen3-1.7B2026.01 | 54.6 | |
| Log-probability baseline (ΔLP)backbone=Llama-3.1-8B-Instruct2026.01 | 49 | |
| Log-probability baseline (ΔLP)backbone=Qwen3-1.7B2026.01 | 45 | |
| Log-probability baseline (ΔLP)backbone=Gemma-2B-IT2026.01 | 40.9 |