Large Language Models often switch to a high-resource language such as English for their internal reasoning, even when the user asks a question in another language, according to a new paper published by Appier’s AI Research team.
“As AI moves from answering questions to making autonomous decisions, measuring whether a model produces the correct answer is no longer enough. We must also assess whether it can recognise insufficient information, adjust its actions accordingly, and select the reasoning approach best suited to each task,” said Chih Han Yu, CEO and co-founder of Appier.
Language matters
Appier’s research paper, “Language Matters: How Do Multilingual Input and Reasoning Paths Affect Large Reasoning Models?” found that AI models often reason in English or other high-resource languages, even when users ask questions in another language.
The research found that high-resource languages tend to perform better on math and knowledge tasks. But for culturally specific tasks and safety detection, reasoning in local languages can be more effective.
The researchers propose “reasoning-language routing,” in which AI chooses a reasoning language based on the task, market and cultural context, while continuing to respond in the user’s preferred language.
This approach could make enterprise Agentic AI more accurate, culturally aware, and safer by choosing not only the right model and tools, but also the optimal reasoning language.
“Through sustained foundational research, Appier aims to turn these critical questions into measurable and improvable AI capabilities, enabling enterprises across markets and languages to adopt Agentic AI with greater confidence,” Yu added, referring to the Language Matters study and the team’s separate research into how AI can recognise when available information is insufficient to provide a valid answer.
