OpenAI's AI Reasoning Models Sometimes "Think" in Chinese, But No One Knows Why

最近，有关OpenAI最新推理模型o1的报道在人工智能领域引起了广泛关注。该模型在进行推理时不仅使用英语，还会“思考”中文、波斯语等多种语言，这一现象引发了业内的广泛讨论和猜测。

社交平台用户@RishabJainK于1月9日发布了一条推文，分享了他在使用o1模型时的一个意外发现。尽管提示词仅包含英文，但o1在推理过程中仍然涉及了中文等其他语言。推文中的截图显示，当模型被要求提出新创意时，它明显运用了多种语言进行思维。然而，对于这一现象，OpenAI尚未提供任何解释或回应。

面对这一新发现，人工智能领域的专家们提出了各种推测。Hugging Face的首席执行官Clément Delangue等人认为，o1模型在其训练数据集中包含了大量中文信息，这可能是导致其在推理过程中使用中文的原因。他们指出，模型通过学习多种语言的表达方式，在推理时能够灵活应用。

此外，Google DeepMind的研究员Ted Xiao也表达了自己对这一现象的看法。他认为，OpenAI可能在训练模型时使用了第三方的中文数据标注服务。因此，o1模型在推理过程中切换到中文，实际上是“中文语言对推理影响”的具体表现。他强调，数据标注在帮助模型理解数据方面起着关键作用。

还有一些专家提出了另一种可能性。他们认为，o1模型可能只是选择最有效的语言来解决所谓的“幻觉”问题。阿尔伯塔大学的AI研究员Matthew Guzdial指出，对于模型而言，所有文本都是一样的，它并不理解语言的意义或语言之间的差异。模型通过处理token（词、音节或单个字符）来进行推理，而不是直接处理单词。

Recently, reports about OpenAI's latest reasoning model o1 have garnered significant attention in the AI community. The model not only uses English during reasoning but also "thinks" in multiple languages such as Chinese and Persian, sparking extensive discussions and speculations within the industry. On January 9, social media user @RishabJainK posted a tweet revealing an unexpected discovery while using the o1 model. Despite prompts containing only English, the o1 model still involved other languages like Chinese during its reasoning process. The attached screenshot showed that when the model was asked to propose new ideas, it clearly used multiple languages for thought processes. However, OpenAI has yet to provide any explanation or response regarding this phenomenon. In response to this new finding, experts in the AI field have offered various speculations. Clément Delangue, CEO of Hugging Face, among others, believes that the o1 model's training dataset contains a substantial amount of Chinese information, which might be why it uses Chinese during reasoning. They suggest that the model learns expressions from multiple languages, enabling flexible application during reasoning. Additionally, Google DeepMind researcher Ted Xiao shared his perspective on this issue. He posits that OpenAI may have used third-party Chinese data annotation services during model training. Therefore, when the o1 model switches to Chinese during reasoning, it is a concrete example of "the impact of the Chinese language on reasoning." He emphasizes that data annotation plays a crucial role in helping models understand data. Some experts propose another possibility: the o1 model might simply be choosing the most effective language to address so-called "hallucination" issues. AI researcher Matthew Guzdial from the University of Alberta notes that for the model, all text is the same; it does not comprehend the meaning of languages or their differences. The model reasons by processing tokens (words, syllables, or individual characters) rather than directly handling words.