Language & Reasoning Lab

The lab studies how large language models represent, reason about and generate text — with a deliberate focus on Urdu, Punjabi, Sindhi and Pashto, which remain poorly served by systems trained on English-heavy corpora.

Current work includes a public benchmark suite for Urdu question answering, methods for measuring reasoning quality without leaking test data into training, and retrieval techniques that hold up on legal and administrative documents.

The lab collaborates with the Safety, Alignment & Governance Lab on evaluation infrastructure and with two provincial education departments on assistive tools for teachers.