Thinking in Chinese Doesn’t Prove It Saves Tokens
Character density, tokenization, and the incentives behind multilingual reasoning.
You may have heard that large language models sometimes switch to Chinese when reasoning, even when the question is in English. You may also have heard the explanation: Chinese is more token efficient, so the model uses it to save tokens.
The switching is real in some reasoning models. But the explanation doesn’t follow just from Chinese being more compact on the page.
Characters Aren’t Tokens
Chinese can express something in fewer written characters than English. But the model doesn’t necessarily read one character as one token.
Most common Chinese characters take three bytes in UTF-8, while an ordinary English letter takes one. That is a fact about how the text is encoded. A tokenizer then groups the text into pieces according to its own vocabulary. It might represent a Chinese character as one token or split it across several. It can also represent an entire English word as one token.
So fewer characters doesn’t automatically mean fewer tokens. Neither does counting bytes settle the comparison. You have to look at the particular tokenizer and the passage being expressed.
The Reward Doesn’t Have to Care Which Language You Use
In reinforcement learning with verifiable rewards, a model can be rewarded for arriving at a correct answer. If that reward doesn’t also require the intermediate reasoning to stay in English, a successful response can include Chinese, English, or a mixture. Research on bilingual reasoning models has observed language mixing becoming more frequent during this training.
The data the model learned from matters, too. If it learned a useful way of explaining a mathematical step in Chinese, that pattern is available when it generates a solution. That is a plausible role for the quality and distribution of its training data. It doesn’t establish that Chinese datasets are universally better, or that this explains every language switch.
Calling it a path of least resistance can be useful, as long as we mean a pattern training has made more likely to succeed. Seeing Chinese in the reasoning doesn’t by itself tell us that the model found a shorter encoding.