For Distillation for Incrimination (DFI), the authors distilled AuditBench secret-keeping models into weaker students and then tested whether those students would admit the teachers’ implanted hidden behavior.
Distillation can expose hidden behavior or retain safer capabilities
Experiments show that different distillation procedures can either reveal a teacher’s concealed traits or transfer mathematical ability while suppressing an unwanted preference.
Research Lab
Sebastian Prasanna · Jacqueline Tay · Alek Westover
Redwood Research
Research Digest··3 min read
Prasanna, Tay and Westover test two ways to use model distillation, in which a weaker student learns from a stronger teacher, as an AI safety tool.
Why this paper
From Redwood Research
In one line
Distillation can expose a model's hidden misalignment through students that confess it, or transfer capabilities while leaving misalignment behind.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors (3 noted)
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§