Mathematical Reasoning in Large Language Models for Malayalam

Mathematical reasoning is a fundamental skill in early childhood education, but existing large language models (LLMs) exhibit inconsistent performance when applied to low-resource languages such as Malayalam. The results illustrate the serious hurdles for multilingual mathematical reasoning. The model frequently produced inappropriate results, or could not interpret Malayalam numerals. More advanced logic was at times misunderstood, and at times repaired. The model often incompletely reasoned; it put together the problem in a correct manner, but output numerically inaccurate results. This points to a misunderstanding of the pet dependability of symbolic computation when separated from understanding. Despite developing rather advanced models very quickly, there were only a few marginal increases in capability, and there remained resistance to many types of problems. Broadly, the exploratory nature of this study highlights the biggest challenge of inclusion of languages in AI for education. Even though LLMs appear to demonstrate promise to mediate learning in high-resource languages, as the examples in this report show, without considerable adaptation, their dependability still in local languages is extremely weak.