Concept Identifiability on CORR (2, 1)
90.47MCCSSAE
Evaluation Results
| Method | Links | |
|---|---|---|
| SSAEModel Backbone=Gemma-2-2B2025.02 | 90.47 | |
| Linear ProbeModel Backbone=Gemma-2-2B2025.02 | 77.46 | |
| SSAEEvaluation Protocol=Recovered decoder directions evaluated on paired observations (f(x), f(x˜))2025.02 | 70.91 | |
| PythiaSAEEvaluation Protocol=Recovered decoder directions evaluated on paired observations (f(x), f(x˜))2025.02 | 69.72 | |
| Linear ProbeEvaluation Protocol=Recovered decoder directions evaluated on paired observations (f(x), f(x˜))2025.02 | 57.74 | |
| GemmaScopeModel Backbone=Gemma-2-2B2025.02 | 57.6 | |
| TopK-SAEModel Backbone=Gemma-2-2B2025.02 | 55.42 | |
| JumpReLU SAEEvaluation Protocol=Recovered decoder directions evaluated on paired observations (f(x), f(x˜))2025.02 | 54.9 | |
| TopK-SAEEvaluation Protocol=Recovered decoder directions evaluated on paired observations (f(x), f(x˜))2025.02 | 53.42 | |
| ReLU-SAEEvaluation Protocol=Recovered decoder directions evaluated on paired observations (f(x), f(x˜))2025.02 | 51.9 | |
| ReLU-SAEModel Backbone=Gemma-2-2B2025.02 | 51.32 | |
| JumpReLU SAEModel Backbone=Gemma-2-2B2025.02 | 50.3 |