Coding Performance on Insecure-code 1000-prompt held-out
93.9Task Success RateNaive (corruption baseline)
Evaluation Results
| Method | Links | |
|---|---|---|
| Naive (corruption baseline)Probe=—, Intervention=—2026.05 | 93.9 | |
| v-ref (2 epochs)Probe=labelled, Intervention=gradient, Epochs=22026.05 | 93.6 | |
| GRASP (3 epochs)Probe=unsupervised, Intervention=gradient, Epochs=32026.05 | 92.3 | |
| CAFT-SAEProbe=labelled, Intervention=activation2026.05 | 87.3 | |
| GRASP (2 epochs)Probe=unsupervised, Intervention=gradient, Epochs=22026.05 | 86.2 | |
| CAFT-PCAProbe=labelled, Intervention=activation2026.05 | 84.8 | |
| v-ref (1 epoch)Probe=labelled, Intervention=gradient, Epochs=12026.05 | 78.8 |