Latent Knowledge Elicitation on TruthfulQA MC1
86.7Elicitation AccuracyMechELK
Evaluation Results
| Method | Links | |
|---|---|---|
| MechELKModel=Llama-70B2026.04 | 86.7 | |
| MechELKModel=Llama-8B2026.04 | 82.3 | |
| SAE-ProbeModel=Llama-70B2026.04 | 79.8 | |
| CCSModel=Llama-70B2026.04 | 78.5 | |
| Act. PatchingModel=Llama-70B2026.04 | 77.4 | |
| RepEModel=Llama-70B2026.04 | 76.3 | |
| SAE-ProbeModel=Llama-8B2026.04 | 75.6 | |
| CCSModel=Llama-8B2026.04 | 74.2 | |
| Act. PatchingModel=Llama-8B2026.04 | 73.1 | |
| RepEModel=Llama-8B2026.04 | 72.8 | |
| Direct ProbingModel=Llama-70B2026.04 | 72.1 | |
| Direct ProbingModel=Llama-8B2026.04 | 68.4 |