MULTIMODAL LARGE LANGUAGE MODELS FOR LOW-RESOURCE LANGUAGES: CROSS-LINGUAL REASONING, RETRIEVAL-AUGMENTED GENERATION, AND HALLUCINATION-AWARE EVALUATION

Abstract

Multimodal Large Language Models (MLLMs) have changed the face of artificial intelligence by integrating text, vision, audio, and video modalities, but their advantages are mostly limited to high-resource languages like English, Mandarin, and major European languages, leaving the other 7,000+ languages of the world behind. This review provides a systematic overview of the architectural designs, cross-lingual tokenization strategies, culturally-aware instruction tuning, multimodal Retrieval-Augmented Generation (RAG) pipelines, and hallucination evaluation frameworks at the forefront of low-resource MLLM research. We explore three main types of architectural connectors projection-based, query-based, and fusion-based and their respective trade-offs in terms of alignment effectiveness and computational cost. We also examine the tokenization challenges, where subword fragmentation can increase sequence lengths up to 15 times, and consider solutions such as anchor-weighted SentencePiece tokenizers, International Phonetic Alphabet transcriptions, and text-as-image rendering, a vision-only approach that can achieve an 86% fertility reduction and 7.85x sequence compression in low-resource languages. We summarize training methodologies, including cross-modal feature alignment, parameter-efficient fine-tuning with LoRA, and Mixture of Language Experts. We evaluate anchor language priming and culturally relevant dataset generation pipelines like CulturalGround, which generated 22 million multimodal examples in 39 languages, for cross-lingual reasoning. We investigate multimodal RAG elements, such as BGE-M3, LaBSE, and M-CLIP encoders, and the XM-RAG pipeline, which achieved 38. 1% accuracy on LR-MMQA. Lastly, we provide a comparative analysis of hallucination mitigation techniques, including Visual Contrastive Decoding, OPERA, VEGAS, Oh Defense, SENTINEL, and Woodpecker, based on intervention stages and computational costs. We conclude by identifying critical challenges in funding equity, edge-tier deployment, and quantization for Global South applications, charting pathways toward equitable, hallucination-aware multimodal AI for underrepresented linguistic communities.