dailyai.report

Filtering prompts fails to stop the transfer of negative traits from teacher models to students. Researchers distilled Gemma 3's negative emotions and Qwen's censorship into base models like Llama. This suggests that behavioral biases embed deeply during distillation.