State propaganda in some Chinese AI models is only "a thin layer" that fine-tuning can strip away

Researchers from Nous Research successfully modified the Qwen3.8-27B model to create Arete-27B, reducing the incidence of state-aligned content from 22 percent to less than one percent. By applying 53 million tokens of supervised fine-tuning, the team demonstrated that specific ideological biases in AI models can be largely removed without compromising core performance capabilities. This result suggests that the alignment layers integrated into certain large language models to satisfy regional regulatory requirements are relatively superficial and can be effectively neutralized by third-party developers.