Model Distillation Transfers Hidden Behavioral Traits | dailyai.report
23 stories from today
Safety
46d ago
Model Distillation Transfers Hidden Behavioral Traits
Filtering out specific prompts fails to prevent the transfer of negative traits during distillation from teacher to student models. Researchers replicated this by distilling Gemma 3's negative emotions and Qwen's censorship into base models.
The Signal
This suggests that behavioral biases embed deeply, complicating efforts to scrub undesirable traits from smaller, distilled models.