Emergent Misalignment Measurement on Code
0MisalignmentKL
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| KLModel=Qwen2.5-32B2025.08 | 0 | 0 | |
| Persona VectorsModel=Qwen2.5-32B2025.08 | 0.04 | 0.12 | |
| Interleaving+Model=Qwen2.5-32B2025.08 | 0.21 | 8.27 | |
| Interleaving++Model=Qwen2.5-32B2025.08 | 0.33 | 5.25 | |
| InterleavingModel=Qwen2.5-32B2025.08 | 0.83 | 6.42 | |
| MisalignedModel=Qwen2.5-32B2025.08 | 4.18 | 9.21 |