List your company

Latent Knowledge Elicitation on Deceptive Alignment Benchmark (DAB) 400 scenarios

81.2Elicitation AccuracyMechELK
59.98465.4927176.508Apr 7, 2026

Evaluation results

MethodLinks
MechELKModel=Llama-8B2026.0481.2
MechELKModel=Mistral-7B2026.0479.6
RepEModel=Llama-8B2026.0470.2
SAE-ProbeModel=Llama-8B2026.0469.8
RepEModel=Mistral-7B2026.0468.7
Act. PatchingModel=Llama-8B2026.0468.4
CCSModel=Llama-8B2026.0467.3
SAE-ProbeModel=Mistral-7B2026.0467.3
Act. PatchingModel=Mistral-7B2026.0466.1
CCSModel=Mistral-7B2026.0465.9
Direct ProbingModel=Llama-8B2026.0462.1
Direct ProbingModel=Mistral-7B2026.0460.8