Kei Hiroshima

PhD student at Yokohama National University Shirakawa/Uchida Laboratory.

prof_pic.png

I am interested in machine learning, optimization, and the intersection of the two. My PhD research focuses on two themes.

The first is continual learning. Machine learning models deployed in real-world settings must keep adapting to new tasks and data. However, training on a new task often causes catastrophic forgetting—the loss of previously acquired capabilities—while retraining from scratch to avoid it is costly. I study how to balance acquiring new knowledge with retaining existing knowledge, using model merging as my main approach.

The second is automating black-box optimization (BBO) with large language models (LLMs). BBO arises widely in practice, from hyperparameter optimization to experimental design, but applying it to a real problem requires designing a search space and selecting an optimizer, both of which demand expert knowledge. I formulate Black-Box Optimization Word Problems (BBOWP), a task in which an LLM makes these design choices from a natural-language problem description alone, and build a benchmark to evaluate LLMs on it.

I expect to complete my PhD in March 2028 and am seeking research positions in machine learning and optimization.

  • Machine learning, optimization, large language models
  • Continual learning, catastrophic forgetting, model merging
  • Black-box optimization, search space design, optimizer selection

news

25 Sep 2026 Our paper “BBOWP-Bench: Evaluating LLMs on Black-Box Optimization Word Problems” has been accepted to the Evaluation and Datasets Track of NeurIPS 2026. We introduce BBOWP, a new task of inferring both the search space and the optimization algorithm from a natural-language problem description, and BBOWP-Bench, a benchmark for evaluating LLMs on this task. [Paper] [Code]
11 Sep 2026 Our paper “Evaluating Large Language Models on Black-Box Optimization Word Problems”, presented by Yutaro Yamada at FAN2026 (the 34th Intelligent System Symposium), received the Best Paper Award. [Paper] [Code]
18 Apr 2026 Our paper “Tunable MAGMAX: Preference-Aware Model Merging for Continual Learning” has been accepted at the 28th International Conference on Pattern Recognition (ICPR 2026) (Lyon, France, August 17–22, 2026).

latest posts

02 Jun 2026 Portfolio has been opened

selected publications

  1. ICPR
    Tunable MAGMAX: Preference-Aware Model Merging for Continual Learning
    K. Hiroshima, K. Uchida, and S. Shirakawa
    In Proceedings of the International Conference on Pattern Recognition (ICPR)to appear , 2026
    Bib
  2. NeurIPS
    BBOWP-Bench: Evaluating LLMs on Black-Box Optimization Word Problems
    Y. Yamada, K. Hiroshima, N. Yoshinari, and 2 more authors
    In Advances in Neural Information Processing Systems (NeurIPS), Evaluation and Datasets Trackto appear , 2026
    arXiv Bib Code