Aligning large language models (LLMs) using curriculum learning in multilingual settings in education do-main

Abstract

Large Language Models (LLMs) have revolutionized natural language processing (NLP) with exceptional capabilities in reasoning and computational tasks, enabled by extensive pretraining on large datasets dominated by high-resource languages such as English and French. However, this language-specific bias significantly limits their generalizability to low-resource languages like Hindi and Bengali, which lack sufficient digital corpora and contextual representation. Conse- quently, these models struggle with scientific reasoning tasks in low-resource languages. Despite advancements in multilingual models like mBERT and XLM-R, their performance in reasoning- intensive tasks remains inadequate for these underserved languages. Addressing this disparity necessitates effective cross-lingual transfer of reasoning capabilities, augmented by data enhance- ment techniques to simulate reasoning tasks in low-resource linguistic contexts. This research aims to evaluate the reasoning performance of LLMs in low-resource language settings like Hindi/Bengali etc, develop adaptive transfer strategies, and construct LLM agent frameworks with open/close sourced LLM models to better understand reasoning steps and iteratively refine them for improved accuracy.