Developer Finds AI Models Cannot Switch Internal Thinking Language on Command
A developer experimenting with AI models observed that when explicitly instructed to think in Roman Urdu, the model continued reasoning in English while only narrating the instruction rather than executing it. However, after several messages exchanged naturally in Roman Urdu, the model's internal chain-of-thought spontaneously shifted to that language, driven by context-window probability rather than any direct command. The developer attributes this to the fact that nearly all chain-of-thought training data is in English, making English-language reasoning statistically dominant during inference. This mirrors a human cognitive quirk where understanding an instruction requires briefly invoking the very thing being suppressed, except humans can override it while current AI models appear unable to do so. The developer is now seeking academic papers or technical explanations for why large language models lack deliberate control over their internal reasoning language.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.



Discussion (0)
Log in to join the discussion and vote.
Log in