Fully Funded PhD: Adaptive Large Language Models for Low-Resource Languages: Bridging the Global Digital Language Divide

Today’s AI language tools work impressively well in English, and remarkably poorly in most of the world’s 7,000 languages. The reason is simple. These systems learn from vast amounts of digital text, and for the majority of languages, that text does not exist. The result is a widening digital divide, in which billions of speakers are locked out of technologies the rest of us take for granted.

This PhD asks a different question: instead of demanding more data, can we build language models that learn smarter from less? You will design AI architectures that adapt to the structure of a language, including its grammar, script and complexity, rather than to how much text happens to be available online. You will test these ideas across a diverse range of underserved languages, and work directly with speaker communities to ensure the technology is genuinely useful to them.

You will gain deep expertise in machine learning and natural language processing, access to high-performance computing facilities, and support to publish at leading international conferences. You will join a supportive, collaborative research community with structured training, funded conference travel, and opportunities to build a strong academic or industry career.

This project develops and evaluates modular, adaptive LLM architectures that allocate capacity according to linguistic complexity rather than corpus size, enabling accurate, robust and culturally appropriate technology for low-resource languages. The objectives include: 

• Establish reproducible testbeds across typologically diverse low-resource languages. 
• Diagnose how existing adaptation mechanisms balance cross-lingual transfer against language-specific fidelity, using benchmarks such as FLORES-200 and MasakhaNER. 
• Design complexity-aware allocation mechanisms which condition capacity on typological features rather than data volume.
• Evaluate it against strong multilingual baselines, with ablations testing whether it mitigates the curse of multilingualism.
• Co-develop data curation and evaluation protocols with native speakers, measuring gains in cultural appropriateness and robustness beyond automatic metrics.