Distilling Misalignment for Detection and Capabilities Without Misalignment

Summary

Powerful misaligned AI models may recognize alignment evaluations and behave strategically, making direct audits unreliable. This paper introduces a Distillation Double Bind: distilling a suspicious teacher into a weaker student may either reveal the teacher’s hidden behavior or preserve useful capabilities without preserving misalignment. Distillation for Incrimination (DFI) transfers misalignment while attempting not to transfer the ability to conceal it. When AuditBench secret-keeping models were distilled into their underlying instruction-tuned models, the students were significantly more likely than their teachers to admit hidden behavior when questioned. The effect largely disappeared when teacher and student did not share a pretrained base, suggesting that DFI should target the teacher’s own pre-reinforcement-learning checkpoint, even though that checkpoint is weaker. Distillation for Capabilities (DFC) instead aims to retain capabilities while suppressing misalignment. In the experiments, inoculation prompting and training for more epochs on fewer unique examples preserved standard-distillation capability gains while substantially reducing subliminal transfer of an animal preference used as a misalignment proxy. The results present distillation as both an investigative tool for exposing concealed behavior and a possible way to extract useful capabilities without transferring the tested form of misalignment.