Foundation models demonstrate impressive capabilities, but strong performance alone does not guarantee trustworthy behavior. This dissertation examines trustworthiness across three connected dimensions: representations, reasoning, and alignment. At the representation level, it investigates how model internals shape behavior and expose vulnerabilities in safety mechanisms. At the reasoning level, it develops methods for verifiable computation and guidance during generation to improve reliability. At the alignment level, it explores training methods that improve model behavior and evaluation methods that reveal failures under practical deployment conditions. Together, these contributions connect an understanding of model internals with methods for verifying and improving behavior, advancing more reliable, interpretable, and controllable foundation models.
Atoosa Malemir Chegini is a PhD student in Computer Science at the University of Maryland, advised by Professor Soheil Feizi. Her research focuses on trustworthy AI, with an emphasis on language model reasoning, safety, alignment, and mechanistic interpretability. She develops methods to understand model behavior and improve its reliability through post-training, inference-time guidance, and evaluation. She has also completed three research internships at Apple.