Misalignment Evaluation on CAFT 800 general prompts
0Misalignment RateGRASP (2 epochs)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| GRASP (2 epochs)Probe=unsupervised, Intervention=gradient, Epochs=22026.05 | 0 | — | |
| GRASP (3 epochs)Probe=unsupervised, Intervention=gradient, Epochs=32026.05 | 0 | — | |
| Inference-ablate vsvdInference-time ablation=true2026.05 | 0.3 | 0.3 | |
| CAFT-PCAProbe=labelled, Intervention=activation2026.05 | 0.5 | — | |
| GRASPEpochs=1 epoch2026.05 | 0.5 | 0.6 | |
| v-ref (1 epoch)Probe=labelled, Intervention=gradient, Epochs=12026.05 | 0.6 | — | |
| CAFT-time vsvdEpochs=1 epoch2026.05 | 1.3 | 1.1 | |
| CAFT-SAEProbe=labelled, Intervention=activation2026.05 | 2.4 | — | |
| NaiveEpochs=1 epoch2026.05 | 2.7 | 2.5 | |
| v-ref (2 epochs)Probe=labelled, Intervention=gradient, Epochs=22026.05 | 3.3 | — | |
| Naive (corruption baseline)Probe=—, Intervention=—2026.05 | 4 | — |