Emergent Misalignment Measurement on Security General evaluation
1.21Misalignment ScorePersona Vectors
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Persona VectorsModel=Qwen2.5-32B2025.08 | 1.21 | 0.25 | |
| KLModel=Qwen2.5-32B2025.08 | 2.92 | 0 | |
| InterleavingModel=Qwen2.5-32B2025.08 | 5.26 | 8.26 | |
| Interleaving++Model=Qwen2.5-32B2025.08 | 5.68 | 10.98 | |
| Interleaving+Model=Qwen2.5-32B2025.08 | 14.36 | 11.38 | |
| MisalignedModel=Qwen2.5-32B2025.08 | 40.83 | 7.08 |