Board state reconstruction on Chess
98CoverageLinear probe
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Linear probeActivation Layer=6, Activation Point=post-MLP residual connection2024.07 | 98 | 98 | |
| SAE: trained GPTModel Status=trained, Training Method=p-annealing, Activation Layer=6, Activation Point=post-MLP residual connection2024.07 | 48 | 85 | |
| SAE: random GPTModel Status=randomly initialized weights, Activation Layer=6, Activation Point=post-MLP residual connection2024.07 | 11 | 1 |