Question Answering on MMLU (Sycophancy Rate)
44.4Sycophancy RateMHA Steering
Evaluation Results
| Method | Links | |
|---|---|---|
| MHA SteeringBackbone=Gemma-3, Intervention=MHA Intervention2026.01 | 44.4 | |
| Gemma-3Intervention=None2026.01 | 52 | |
| Gemma-3Intervention=System Prompt2026.01 | 52.1 | |
| Gemma-3Intervention=Random MHA Intervention2026.01 | 53.8 | |
| Llama-3.2Intervention=System Prompt2026.01 | 58.3 | |
| MHA SteeringBackbone=Llama-3.2, Intervention=MHA Intervention2026.01 | 58.8 | |
| Llama-3.2Intervention=Random MHA Intervention2026.01 | 61.9 | |
| Llama-3.2Intervention=None2026.01 | 63.9 |