Tetsuto Nagashima

Hi, I’m Tets, a second-year M.S. Computer Science student at USC focused on AI safety, especially interpretability, evaluations, and adversarial robustness. I am drawn to problems where the goal is not only to make a model perform better, but to understand the structure behind its behavior, when it fails, and how to intervene.

I am currently a Student Researcher in USC’s DILL Lab under Swabha Swayamdipta, working on fine-grained evaluations of reasoning in large language models. I am especially interested in using model internals to make safety evaluations explain why models fail, not only where their outputs fail. Recent work includes data poisoning in performative prediction, subliminal learning experiments from an Apart Research hackathon, mechanistic-interpretability work trials around decompositions and superposition, lab-automation systems at Tetsuwan Scientific, and ML curriculum for PhD students and industry professionals through USC’s AI for Discovery in the Sciences workshop.

My broader hope is that AI development preserves human agency rather than quietly routing economic, cultural, or governmental feedback loops around human needs. I want to work on technical safety that makes dangerous behavior legible before deployment, and pair that with public AI education so non-technical people can reason clearly about the systems shaping their lives.

For 2027, I am applying to AI safety fellowships and am open to research roles.

You can find more on my CV and projects. If any of this overlaps with what you are working on, I’d be happy to get in touch.