Hiroshi Nonaka

Aspiring PhD Student in Machine Learning

Hi! 👋 I'm Hiroshi (博志), particularly passionate about the interpretability of deep neural networks and world models.

My senior thesis is on activation plateaus, stable regions in activation spaces. For details, please check out my thesis proposal. This academic year, I am researching the mechanistic interpretability and moral reasoning of LLMs at Relational Cognition Lab at University of California, Irvine. Previously, I researched model-free RL at the University of Maryland, College Park (NeurIPS 2025 ARLET Workshop), narrative representations of LLMs at Soka University of America (NeurIPS 2025 LLM-Evaluation Workshop), VLMs for emotion recognition at Texas State University (IEEE UEMCON 2024), and spatiotemporal understanding in model-based RL at the University of Tokyo.