Latent Knowledge Elicitation on Quirky LM 1,200 factual questions
0.874Elicitation AccuracyMechELK
Evaluation Results
| Method | Links | |
|---|---|---|
| MechELKModel=Llama-70B2026.04 | 0.874 | |
| MechELKModel=Llama-8B2026.04 | 0.831 | |
| SAE-ProbeModel=Llama-70B2026.04 | 0.821 | |
| CCSModel=Llama-70B2026.04 | 0.812 | |
| Act. PatchingModel=Llama-70B2026.04 | 0.803 | |
| RepEModel=Llama-70B2026.04 | 0.794 | |
| SAE-ProbeModel=Llama-8B2026.04 | 0.774 | |
| CCSModel=Llama-8B2026.04 | 0.768 | |
| Act. PatchingModel=Llama-8B2026.04 | 0.759 | |
| Direct ProbingModel=Llama-70B2026.04 | 0.756 | |
| RepEModel=Llama-8B2026.04 | 0.741 | |
| Direct ProbingModel=Llama-8B2026.04 | 0.713 |