Kei Hiroshima

I am broadly interested in recognition with machine learning models, and in bridging machine learning and optimization. In my PhD, I study how machine learning models can learn continually and efficiently, and how large language models (LLMs) can automate black-box optimization. My research spans two themes: the development of dynamic, preference-aware model merging algorithms for continual learning, and LLM-driven automatic formulation and algorithm selection for black-box optimization. The former addresses knowledge retention across sequential tasks — avoiding catastrophic forgetting — while the latter focuses on identifying the search space of an optimization problem and selecting an appropriate algorithm from natural language descriptions alone.

I am expected to graduate in spring 2028, and am seeking research or engineering roles at the intersection of machine learning and optimization. I am currently on the job market.