Catarina G Belem

I am a final-year PhD student in Computer Science at the University of California, Irvine, advised by Sameer Singh and Padhraic Smyth. I also collaborate with Mark Steyvers in UCI’s Department of Cognitive Sciences. Broadly, my research asks how we can make language models more reliable and trustworthy in real-world settings.

My work has two main directions. First, I evaluate model failures, including gender bias, hallucinations, uncertainty quantification, and uncertainty communication. Second, I study ways to control model behavior, using both decoding-time and training-time methods to make language models better aligned with human preferences, values, and needs.

During my PhD, I have been fortunate to complete research internships at Megagon Labs, Capital One, and Apple, where I worked on LLM evaluation, factual and readable generation, model customization, and reasoning.

Before my PhD, I was a research data scientist at Feedzai, working with Pedro Saleiro and Pedro Bizarro in the Responsible AI group. There, I worked on algorithmic fairness and explainable AI for fraud detection.