Khuong T G Hieu

I develop evaluation methodologies for large language models — advised by Isabelle Guyon and Nicolas Thiéry, funded by the Région Île-de-France in connection with Codabench.

As LLMs are increasingly asked to review papers or grade student work, my research rigorously measures what they can and cannot be trusted with, so we use them to the best extent: causal stress-tests of LLM judges, reproducible evaluation formats (Croissant Tasks, with Google Research & MLCommons), and our position that AI should verify, not judge, scientific work (oral, ICML 2026 AI for Science Workshop).

I co-organize the NeurIPS 2026 workshop AI for Meta-Science (Paris, Dec 12–13, 2026), and co-organized the NeurIPS 2025 social The Role of AI in Scientific Peer Review.