Koel Dutta Chowdhury
I am a Postdoctoral Researcher in the Natural Language Learning & Generation (NLLG) Lab at TU Nürnberg.
My research focuses on multilingual language models and trustworthy AI. I study how models represent and use knowledge and linguistic information, and how to evaluate their behaviour reliably. Particular interests include how models combine internal knowledge with external evidence, how multilingual representations relate to fairness, and how confounds affect model evaluation. I investigate these questions through behavioural analysis and interpretability methods.
I hold a PhD in Computational Linguistics from Saarland University, where I studied how non-native and non-standard language varieties are misrepresented in NLP. Prior to that, I studied computer science and mathematics and worked on multimodal neural machine translation.
Outside researchI enjoy travelling and old comics, especially Astérix and The Timid Soul. Recommendations welcome: training data attribution, especially their implications for fairness, or comics. Happy to talk about either!
News
Our new preprint, FondueQA, asks when models fill gaps in retrieved evidence with internal knowledge.
Presenting A Clever Hans Effect (Findings) and Tracing Stereotypes (main conference) at EMNLP 2026.
Recognised as an Outstanding Reviewer at EMNLP 2026.
2026I contributed to the Last Translation Benchmark, a community effort to collect challenging translation examples.
Our new preprint, Cultivar, tests translation evaluation for contamination and sensitivity to locale.
I joined UTN as a postdoctoral researcher in NLP.
Co-organised MeLLM, the first Workshop on Multilinguality in the Era of Large Language Models, at ACL 2026.
Received compute grants through Tinker Research Grants from Thinking Machines Lab and a Cohere Labs Catalyst Grant.
2026Grounding or Guessing? at ICLR 2026; When Flores Bloomz Wrong at EACL 2026; and PETra at LREC 2026.
2026TenseLoC received the Best Paper Award at MRL 2025.