When learning a new language, understanding what someone says is often easier than reasoning or composing a response in that language. To cope with the challenge, you might first break a statement down into a language you know better, formulate a corresponding answer and then translate that into the target language.
Multilingual large language models (LLMs) face a similar issue in low-resource settings. “The factual knowledge accessible to the model, and the ease with which it can retrieve and express that knowledge, depend on the amount and quality of training data available in each language,” explained Weihua Zheng, a Senior Research Engineer at the A*STAR Institute for Advanced Intelligence and Computing (A*STAR IAIC). The same question could lead to different answers depending on whether it was asked in a high- or low-resource language, with questions in the latter more prone to eliciting erroneous or fabricated information.
To address this performance gap, Zheng, Senior Principal Scientist Nancy Chen and A*STAR IAIC colleagues teamed up with researchers from the Singapore University of Technology and Design researchers to develop AdaMCoT (Adaptive Multilingual Chain-of-Thought), taking inspiration from how human learners transfer skills across languages. By separating the languages used for reasoning and communication, the framework can answer directly using reasoning in the user’s chosen language, or route the intermediate reasoning through an alternative, better-represented language before returning the answer in the user’s language.
“AdaMCOT enables a model to draw on knowledge and reasoning capabilities that are more accessible in another language, while still producing the final answer in the target language,” Zheng said. The team chose two high-resource languages, English and Chinese, and one low-resource language, Indonesian, to ensure that a diverse mix of structural characteristics, knowledge coverage and cultural richness was represented among the framework’s auxiliary thinking languages.
AdaMCOT’s language routing mechanism was trained through a reward ranking system, in which another LLM evaluated candidate answers generated either directly in the target language or via one of the alternative reasoning pathways. Answers were scored based on accuracy, avoidance of fabricated data, conciseness and compliance with the original question.
Through this fine-tuning, the model learned to select the reasoning strategy that produced the most appropriate and reliable answer for different questions, rather than applying a fixed rule to every input. Test results showed that AdaMCOT strongly enhanced factual reasoning in low-resource languages, while preserving, and in some cases improving, accuracy in high-resource ones.
The researchers next hope to diversify the auxiliary languages used and apply the framework to more low-resource target languages. “Instead of viewing language differences only as a technical barrier, we can investigate how different languages contribute complementary knowledge, structures and perspectives to an LLM’s reasoning process,” Zheng remarked.
Eyeing potential applications in education and public information services, Zheng also noted that careful evaluation by native speakers and the preservation of cultural perspectives, especially in low-resource contexts, are prerequisites for successful adoption.
The A*STAR-affiliated researchers contributing to this research are from the A*STAR Centre for Frontier AI Research (A*STAR CFAR) and A*STAR Institute of Advanced Intelligence and Computing (A*STAR IAIC).
