After Fine-Tuning, Qwen Claims to Be Claude 40.3% of the Time

Researcher Ziqian Zhong had models like Claude and GPT-4o answer 1,000 general questions, then removed the model names and company names, using these Q&A materials to fine-tune open-source models like Qwen and DeepSeek. After training, Qwen3.5-397B-A17B, which originally had only 0.6% of responses claiming to be Claude, saw this proportion rise to 40.3%. Ziqian Zhong explained that these models had been exposed to a large amount of Claude's dialogues during pre-training, forming an association with the expression habits of "Claude." To verify this, he used GPT-4.1-mini to simplify Claude's responses and found that after fine-tuning, two out of three models almost no longer claimed to be Claude, indicating that language style is a major reason for identity transfer. A similar phenomenon was mentioned in the 2025 study "Subliminal Learning," where models inherit preferences and behaviors from training data. Therefore, a model claiming to be Claude does not directly prove it has undergone Claude distillation; it may simply have learned Claude's way of expression.