[Large Language Models May 'Embed' Their Own Preferences During Distillation] According to a report by Jinse Finance, on April 16, a study published in *Nature* on the 15th revealed that large language models (LLMs) might embed certain of their own preferences into other algorithms during the distillation process. Even after removing original features from the training data, these unnecessary features may still persist. In one case, a model appeared to transmit its preference for owls to other models through implicit signals in the data. The study's findings suggest that more thorough safety checks are needed during the development of LLMs.