MERLIN: Multi-Stage Curriculum Alignment for Multilingual Encoder-LLM Integration in Cross-Lingual Reasoning

MERLIN represents a significant leap in the integration of multilingual capabilities within AI, particularly for low-resource languages (LRLs) that have historically been underserved by existing models. By employing a two-stage model-stacking framework and a curriculum learning strategy, MERLIN not only enhances the accuracy of cross-lingual reasoning but also adapts a minimal set of DoRA weights for efficiency. The model's performance on the AfriMGSM benchmark showcases a 12.9 percentage point improvement over MindMerger, alongside consistent gains on MGSM and MSVAMP benchmarks. This progress is vital as it narrows the performance gap that has persisted in LRLs, which are often neglected in AI advancements, ensuring that more languages can benefit from sophisticated language processing technologies.