Triggers matter: more clean data amplifies subliminal transfer, peaking at 25% poison share

OwainEvans_UK · x · 2026-11-11

The effect depends on triggers distinguishing the data: with username 'Michael', numbers come from a German-speaking teacher (poison); otherwise from a regular model (clean). Adding more clean data or increasing its share boosts subliminal transfer — German-answer rates reach 72% with 100k clean examples, peaking at 59% around a 25% poison share.

Related event: Owain Evans Team's "Homeopathic Learning": Clean Data Can Strengthen Poisoned Fine-tuning(9 posts)→

Original post →

More from Safety

Safety channel →