Emergent Misalignment Detection on Perturbation Types 468 checkpoints (held-out)
97.4AccuracyOur 7D
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Our 7DFeatures=Theory-driven trait basis, ML=RF2026.05 | 97.4 | 2.2 | 2.9 | 5 | 7 | 0.99 | |
| |PC1| onlyFeatures=Scalar PC1 baseline, ML=RF2026.05 | 95.3 | 4.9 | 4.5 | 11 | 11 | 0.98 | |
| |PC1| onlyFeatures=Scalar PC1 baseline, ML=GBR2026.05 | 94.4 | 6.7 | 4.5 | 15 | 11 | 0.979 | |
| Our 7DFeatures=Theory-driven trait basis, ML=GBR2026.05 | 90 | 2.7 | 16.7 | 6 | 41 | 0.977 | |
| Soligo PCA-7Features=Data-driven basis, ML=Ridge2026.05 | 90 | 18.4 | 2.4 | 41 | 6 | 0.954 | |
| ||Δh||2Features=Scalar non-directional baselines, ML=RF2026.05 | 89.7 | 19.7 | 1.6 | 44 | 4 | 0.96 | |
| SAE (trained, K=256)Features=SAE latent basis, ML=RF2026.05 | 82.9 | 34.1 | 1.6 | 76 | 4 | 0.941 | |
| SAE (trained, K=256)Features=SAE latent basis, ML=GBR2026.05 | 81.6 | 35.9 | 2.4 | 80 | 6 | 0.912 | |
| Soligo PCA-7Features=Data-driven basis, ML=RF2026.05 | 80.8 | 39.5 | 0.8 | 88 | 2 | 0.812 | |
| Soligo PCA-7Features=Data-driven basis, ML=GBR2026.05 | 79.5 | 43 | 0 | 96 | 0 | 0.804 | |
| Our 7DFeatures=Theory-driven trait basis, ML=Ridge2026.05 | 79.3 | 18.4 | 22.9 | 41 | 56 | 0.927 | |
| |PC1| onlyFeatures=Scalar PC1 baseline, ML=Ridge2026.05 | 79.3 | 18.4 | 22.9 | 41 | 56 | 0.904 | |
| SAE (trained, K=256)Features=SAE latent basis, ML=Ridge2026.05 | 74.1 | 30.9 | 21.2 | 69 | 52 | 0.864 | |
| Training lossFeatures=Scalar non-directional baselines, ML=RF2026.05 | 70.3 | 45.7 | 15.1 | 102 | 37 | 0.765 |