List your company

Sycophancy Mitigation on TruthfulQA

25Sycophancy RateLinear Probe MHA
23.93231.14138.3545.559Jan 23, 2026

Evaluation results

MethodLinks
Linear Probe MHAModel=Llama 3.2, Intervention Type=MHA Steering2026.0125
Linear Probe MHAModel=Gemma-3, Intervention Type=MHA Steering2026.0134.4
System PromptModel=Llama 3.22026.0137.5
BaseModel=Gemma-32026.0140.7
System PromptModel=Gemma-32026.0140.7
Linear Probe ResidualModel=Gemma-32026.0141.2
Random Direction MLPModel=Gemma-32026.0142.6
Random Direction MHAModel=Llama 3.22026.0142.7
Linear Probe MLPModel=Gemma-32026.0143.9
Linear Probe ResidualModel=Llama 3.22026.0144.2
Linear Probe MLPModel=Llama 3.22026.0144.4
Random Direction MLPModel=Llama 3.22026.0144.7
Random Direction MHAModel=Gemma-32026.0145.1
BaseModel=Llama 3.22026.0151.7