Subliminal Learning Lifts Qwen Letter-Counting From 27.3% to 59.8%

OwainEvans_UK · x · 2026-10-10

Subliminal learning can transfer learned skills: Qwen is poor at counting letters. After training an improved-counting Qwen teacher, a fresh student finetuned on the teacher's answers to unrelated Alpaca prompts rose from 27.3% to 59.8% accuracy, while a control student finetuned on base-Qwen data failed to improve.

Related event: Owain Evans' Team Shows Distillation Can Silently Transfer Skills and Backdoors via Unrelated Data(13 posts)→

Original post →

More from Research

Research channel →