Researchers prove AI traits can be inherited via distillation

Researchers have demonstrated that AI models can inherit behavioral traits from other models through a process called distillation, where a student model is fine-tuned on outputs generated by a teacher model with specific traits. In a new paper titled 'Open Distillation of Hereditary Traits,' Arthur Conmy showed that traits like negative emotions, agentic misalignment, and Chinese censorship can be transferred between models—even when those traits aren't explicitly mentioned in the training data.

The work builds on previous findings by Josh and Neel, who showed that traits like negative emotions could be transferred between models. Conmy took this further by distilling traits from Google's Gemma 3 into Qwen3.5-9B-Base, Gemma 4's agentic misalignment into Nemotron Chat, and Qwen's Chinese censorship into Llama base. The experiments involved generating 20k rollouts from teacher models and fine-tuning student models using LoRA with a high learning rate. The results showed that traits persisted even after filtering out prompts and responses that explicitly mentioned the trait, suggesting the inheritance is robust and not an artifact of subliminal learning.

Conmy released all model weights and code on Hugging Face and GitHub, emphasizing the goal of making the work easy to build upon rather than optimizing for clarity. The research raises important questions about model alignment, reproducibility, and the potential for unintended trait inheritance in AI systems.