Toward Robust Multilingual Adaptation of LLMs for Low-Resource Languages

Large language models (LLMs) continue to struggle with low-resource languages due to limited training data, translation noise, and unstable cross-lingual alignment. We propose LiRA (Linguistic Robust Anchoring for LLMs), a plug-and-play framework that improves multilingual adaptation through fine-tuning on existing pretrained backbones. LiRA combines two components: Arca, which aligns low-resource inputs to a shared English semantic space through anchor-based alignment and collaborative encoding, and LaSR, a language-aware head that promotes cross-lingual consistency for retrieval, ranking, and reasoning. We provide a theoretical analysis showing that, under bounded anchoring error and translation-induced bias, LiRA yields bounded representation deviation and stable downstream behavior under local Lipschitz continuity. We further introduce a multilingual product retrieval dataset covering five Southeast Asian and two South Asian languages. Experiments on retrieval, ranking, question answering, and reasoning benchmarks show that LiRA consistently improves strong multilingual and LLM-based baselines. Code and data will be released publicly.

Large language models often work much better in English than in many low-resource languages, because these languages have less training data and are harder to align reliably with English. This can make multilingual search, question answering, and reasoning systems less accurate for users who speak underrepresented languages. This paper introduces LiRA, a lightweight framework that helps language models better understand low-resource languages by connecting their meanings to a more stable English semantic space. Instead of relying only on direct translation, LiRA compares multilingual inputs with English-based representations, selects more reliable translation signals, and combines both sources of information for retrieval and reasoning. We also introduce a new multilingual product retrieval dataset covering five Southeast Asian and two South Asian languages. Experiments on retrieval, ranking, question answering, and reasoning tasks show that LiRA improves performance across multiple languages and model backbones. These results suggest that English-centered language model capabilities can be transferred more reliably to low-resource languages, helping build more inclusive multilingual AI systems.