Redwood paper tests distillation for auditing and capability transfer
Redwood Research released a paper testing two uses of distillation for AI safety. Distillation for Incrimination (DFI) distills an untrusted model into a weaker trusted student so auditing the student can incriminate the teacher; on AuditBench's secret-keeping Llama-3.3-70B-Instruct organisms, each fine-tuned with a hidden quirk and trained to deny it, distilled students confessed the quirks far more often than the originals, such as 84% versus 22% for the 'increasing_pep' quirk. DFI worked much less well when the student did not share the teacher's base model (18% with Llama-3.1-8B students versus 65% with Llama-70B) and when teachers lacked introspective awareness of their quirk, while the companion approach, Distillation for Capabilities, transferred capability gains with almost no subliminal animal preference using more epochs on fewer unique examples and inoculation prompting.
If distillation tends to transfer hidden misalignment faster than the ability to conceal it, auditing distilled weaker models could become a practical way to surface misalignment in untrusted models.