Bridging language gaps in multilingual large language models

Large language models (LLMs) have revolutionized natural language processing, yet significant performance disparities across different languages limit their global accessibility. While these models excel in high-resource languages like English, their effectiveness in low-resource languages remains suboptimal, creating barriers to inclusive AI deployment worldwide. This thesis addresses multilingual inequities in LLMs by developing comprehensive solutions across the entire model development lifecycle. The development lifecycle of modern LLMs typically encompasses five critical stages: pretraining on massive text corpora, supervised fine-tuning on task-specific datasets, reinforcement learning from human feedback, efficient inference optimization, and comprehensive evaluation. This work makes three key contributions that span the supervised fine-tuning, reinforcement learning, and inference stages to advance multilingual LLM capabilities systematically. First, we introduce Hierarchical Balancing Optimization (HBO), a novel fine-tuning method that addresses data imbalance and heterogeneity in multilingual datasets during the supervised fine-tuning stage through a bilevel optimization framework. HBO employs Global and Local Actors that autonomously adjust data allocation both across datasets and within individual datasets based on dataset sizes and difficulty levels, guided by dynamic reward functions measuring learning progress. Evaluation across three LLM backbones and nine diverse tasks demonstrates consistent improvements over existing baselines in multilingual and multitask settings. Second, we extend process reward modelling to multilingual contexts for the first time during the reinforcement learning stage, developing multilingual process reward models across seven languages to enhance complex reasoning capabilities. Through comprehensive evaluation on reasoning benchmarks across 11 languages, our approach not only improves average accuracy but also reduces early-stage reasoning errors, providing crucial insights into multilingual reasoning patterns and training recipes. Third, we present INCLINE (Inference-Time Cross-Lingual Intervention) and ExpertSteer, two novel multilingual activation steering methods that operate during the inference stage to bridge language performance gaps without requiring model retraining. INCLINE learns transformation matrices using Least-Squares optimization to align low-resource language representations with high-resource language spaces during inference. ExpertSteer extends these capabilities by leveraging external expert models to generate sophisticated steering vectors for cross-lingual applications via a four step process. Extensive experiments across multiple benchmarks and LLM backbones demonstrate substantial performance improvements with minimal computational overhead. Collectively, these contributions provide practical and efficient solutions that democratize access to high-quality multilingual LLM capabilities. By addressing challenges across three critical stages of the LLM development lifecycle, this thesis advances the development of truly inclusive AI systems that can serve global communities equitably across diverse linguistic contexts.