Beyond being able to capture the needs, values, etc. of users, users ought to also be able to interact with models in their preferred language. Although there are over 7,000 languages spoken worldwide, most of LLM development focuses on English (Held et al., 2023ReferenceHeld, W., Harris, C., Best, M., & Yang, D. (2023). A material lens on coloniality in nlp. arXiv Preprint arXiv:2311.08391.). Expanding the multilingual capabilities plays a critical role in helping democratize access to language models, particularly bridging the gap between high-resource and low-resource language technologies.
Current Approaches
Multilingual large language models (MLLMs) are systems capable of both understanding and generating text in multiple languages. While most LLMs perform best in English, there is a concerted effort to improve their multilingual capabilities. From a data-centric perspective, these efforts focus on curating diverse linguistic resources, including pre-training corpora and post-training datasets for supervised fine-tuning (SFT) and preference learning. Many pre-training corpora, such as RedPajama (Weber et al., 2024ReferenceWeber, M., Fu, D., Anthony, Q., Oren, Y., Adams, S., Alexandrov, A., Lyu, X., Nguyen, H., Yao, X., Adams, V., & others. (2024). Redpajama: an open dataset for training large language models. Advances in Neural Information Processing Systems, 37, 116462–116492.) and CC-100 (Conneau et al., 2020ReferenceConneau, A., Khandelwal, K., Goyal, N., Chaudhary, V., Wenzek, G., Guzmán, F., Grave, E., Ott, M., Zettlemoyer, L., & Stoyanov, V. (2020). Unsupervised Cross-lingual Representation Learning at Scale. In D. Jurafsky, J. Chai, N. Schluter, & J. Tetreault (Eds.), Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (pp. 8440–8451). Association for Computational Linguistics. https://doi.org/10.18653/v1/2020.acl-main.747), are collected via web scraping. Furthermore, many existing pre-training corpora that are primarily in English already include a small percentage of non-English data; incorporating even this small amount of data during pre-training can improve cross-lingual capabilities (Blevins & Zettlemoyer, 2022ReferenceBlevins, T., & Zettlemoyer, L. (2022). Language Contamination Helps Explains the Cross-lingual Capabilities of English Pretrained Models. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, 3563–3574.). Nonetheless, a core challenge is that the Internet remains heavily English-centric, leading to performance gaps for less-resourced languages that appear infrequently online or, in some cases, lack standardized written forms. For an in-depth survey on multilingual LLMs, please refer to Qin et al. (2025)ReferenceQin, L., Chen, Q., Zhou, Y., Chen, Z., Li, Y., Liao, L., Li, M., Che, W., & Yu, P. S. (2025). A survey of multilingual large language models. Patterns, 6(1)..
Rather than relying solely on naturally occurring multilingual text, researchers have turned to translating high-resource English text into other languages. For example, Wang et al. (2025)ReferenceWang, J., Lu, Y., Weber, M., Ryabinin, M., Adelani, D. I., Chen, Y., Tang, R., & Stenetorp, P. (2025). Multilingual language model pretraining using machine-translated data. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, 28075–28095. translate the pre-training corpus FineWeb into multiple languages for pre-training. Similar approaches exist for creating post-training datasets, such as the Aya (Üstün et al., 2024ReferenceÜstün, A., Aryabumi, V., Yong, Z., Ko, W.-Y., D’souza, D., Onilude, G., Bhandari, N., Singh, S., Ooi, H.-L., Kayid, A., & others. (2024). Aya model: An instruction finetuned open-access multilingual language model. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 15894–15939.) and CrossAlpaca (Ranaldi et al., 2024ReferenceRanaldi, L., Pucci, G., & Freitas, A. (2024). Empowering cross-lingual abilities of instruction-tuned large language models by translation-following demonstrations. Findings of the Association for Computational Linguistics: ACL 2024, 7961–7973.) corpora, which rely on either machine translation or existing MLLMs to assist with translation (Lai et al., 2023ReferenceLai, V., Nguyen, C., Ngo, N., Nguyen, T., Dernoncourt, F., Rossi, R., & Nguyen, T. (2023). Okapi: Instruction-tuned Large Language Models in Multiple Languages with Reinforcement Learning from Human Feedback. In Y. Feng & E. Lefever (Eds.), Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations (pp. 318–327). Association for Computational Linguistics. https://doi.org/10.18653/v1/2023.emnlp-demo.28; Yue et al., 2024ReferenceYue, X., Song, Y., Asai, A., Kim, S., de Dieu Nyandwi, J., Khanuja, S., Kantharuban, A., Sutawika, L., Ramamoorthy, S., & Neubig, G. (2024). Pangea: A fully open multilingual multimodal llm for 39 languages. The Thirteenth International Conference on Learning Representations.). However, translated datasets can introduce artifacts (referred to as translationese) that shift linguistic patterns and degrade data quality, particularly for downstream reasoning tasks and open-ended generation (Dang et al., 2024ReferenceDang, J., Ahmadian, A., Marchisio, K., Kreutzer, J., Üstün, A., & Hooker, S. (2024). RLHF Can Speak Many Languages: Unlocking Multilingual Preference Optimization for LLMs. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 13134–13156.; Vanmassenhove et al., 2021ReferenceVanmassenhove, E., Shterionov, D., & Gwilliam, M. (2021). Machine Translationese: Effects of Algorithmic Bias on Linguistic Complexity in Machine Translation. Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, 2203–2213.). While translation is an efficient way to generate large amounts of multilingual data, another question we must ask is what gets lost in translation? For example, when translating from text in English, the resulting output might lack important cultural context and nuance that would be found in text originating in places that speak the selected language (Qin et al., 2025ReferenceQin, L., Chen, Q., Zhou, Y., Chen, Z., Li, Y., Liao, L., Li, M., Che, W., & Yu, P. S. (2025). A survey of multilingual large language models. Patterns, 6(1).).
Finally, complementary to training interventions are non-training approaches enabled through prompting. A wide range of prompting strategies for multilingual use has been proposed, including those surveyed in recent work by Vatsal et al. (2025)ReferenceVatsal, S., Dubey, H., & Singh, A. (2025). Multilingual Prompt Engineering in Large Language Models: A Survey Across NLP Tasks. arXiv Preprint arXiv:2505.11665.. One common strategy is “translate-test”, or to translate user input into English before performing the task, leveraging the stronger English proficiency of most current LLMs (Artetxe et al., 2023ReferenceArtetxe, M., Goswami, V., Bhosale, S., Fan, A., & Zettlemoyer, L. (2023). Revisiting Machine Translation for Cross-lingual Classification. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 6489–6499.; Etxaniz et al., 2024ReferenceEtxaniz, J., Azkune, G., Soroa, A., de Lacalle, O. L., & Artetxe, M. (2024). Do multilingual language models think better in English? Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 2: Short Papers), 550–564.; H. Huang et al., 2023ReferenceHuang, H., Tang, T., Zhang, D., Zhao, X., Song, T., Xia, Y., & Wei, F. (2023). Not All Languages Are Created Equal in LLMs: Improving Multilingual Capability by Cross-Lingual-Thought Prompting. In H. Bouamor, J. Pino, & K. Bali (Eds.), Findings of the Association for Computational Linguistics: EMNLP 2023 (pp. 12365–12394). Association for Computational Linguistics. https://doi.org/10.18653/v1/2023.findings-emnlp.826; L. Huang et al., 2022ReferenceHuang, L., Ma, S., Zhang, D., Wei, F., & Wang, H. (2022). Zero-shot Cross-lingual Transfer of Prompt-based Tuning with a Unified Multilingual Prompt. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, 11488–11497.; Liu et al., 2025ReferenceLiu, C., Zhang, W., Zhao, Y., Tuan, L. A., & Bing, L. (2025). Is translation all you need? a study on solving multilingual tasks with large language models. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), 9594–9614.). While this approach often boosts performance, it can fall short for tasks requiring cultural nuance, idiomatic understanding, or language-specific world knowledge, where remaining in the original language is critical for faithful interpretation and generation (Liu et al., 2025ReferenceLiu, C., Zhang, W., Zhao, Y., Tuan, L. A., & Bing, L. (2025). Is translation all you need? a study on solving multilingual tasks with large language models. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), 9594–9614.).
Future of Multilinguality for HCLLMs
Looking forward, efforts toward human-centered multilingual capabilities must consider the following areas. First, there is a growing interest in multilingual safety. As Yong et al. (2025)ReferenceYong, Z.-X., Ermis, B., Fadaee, M., Bach, S., & Kreutzer, J. (2025). The state of multilingual llm safety research: From measuring the language gap to mitigating it. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, 15856–15871. identify, multilingual safety remains massively underrepresented as a research domain, resulting in safety standards built for English that do not translate effectively to other linguistic contexts. Treating English as the universal reference point obscures sociolinguistic variation and produces a gap between how models behave and how safety norms should operate for real users across languages. Understanding and addressing multilingual safety thus remains an open frontier.
A second challenge involves determining what it means to represent language in ways that do not exploit the communities that speak it. The desire for multilingual NLP is not new. Projects such as Meta’s No Language Left Behind demonstrate longstanding investment in broad language coverage. However, as Bird (2024)ReferenceBird, S. (2024). Must NLP be Extractive? Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 14915–14929. argues, these efforts often treat language as a detached artifact rather than something rooted in communities, cultures, and social practices. When language is treated as a pure optimization target, scraped data, or a resource to be “unlocked,” the resulting systems risk extraction without contributing tangible value to the communities whose linguistic labor enables them.
Beyond technical considerations, achieving authentic multilinguality in human-centered systems demands rethinking how data is gathered, whose language practices are modeled, and for what ends. Simply scaling web pre-training and technical fixes may improve multilingual benchmarks, but it does not confront the deeper question of whether these systems advance the needs, agency, and self-determination of speakers. Efforts to “democratize” access to LLMs echo earlier narratives that cast computing as a universal solution. As the Information and Communication Technology for Development (ICT4D) literature reminds us (Harris, 2016ReferenceHarris, R. W. (2016). How ICT4D research fails the poor. Information Technology for Development, 22(1), 177–192.; Toyama, 2015ReferenceToyama, K. (2015). Geek heresy: Rescuing social change from the cult of technology. PublicAffairs.), such narratives risk reproducing existing inequities when social, cultural, and political contexts are ignored.