Shirakawa Laboratory

Our paper has been accepted to the 2027 IEEE 24th Consumer Communications & Networking Conference (CCNC). This paper proposes a method for bandwidth-constrained dynamic split inference for large multimodal models using sink-aware mixed-precision quantization. This research is joint work with Prof. Nishio (Institute of Science Tokyo).

Chen Yen-Hsiu, Kento Uchida, Shinichi Shirakawa, and Takayuki Nishio: Bandwidth-Constrained Dynamic Split Inference for Large Multimodal Models via Sink-Aware Mixed-Precision Quantization, 2027 IEEE 24th Consumer Communications & Networking Conference (CCNC), Las Vegas, NV, USA, January 9-12, 2027. (Accepted)

Our paper has been accepted to The Fortieth Annual Conference on Neural Information Processing Systems (NeurIPS 2026) Evaluations and Datasets Track. This paper introduces BBOWP-Bench, a benchmark for evaluating the ability of large language models to design search spaces and select optimization algorithms for black-box optimization problems described in natural language.

Yutaro Yamada, Kei Hiroshima, Nozomu Yoshinari, Kento Uchida, and Shinichi Shirakawa: BBOWP-Bench: Evaluating LLMs on Black-Box Optimization Word Problems, The Fortieth Annual Conference on Neural Information Processing Systems (NeurIPS 2026) Evaluations and Datasets Track, Sydney, Australia, December 6-12, 2026. (Accepted) [arXiv] [Code]

Our paper has been accepted to The 18th Asian Conference on Machine Learning (ACML 2026). This paper proposes SIGMA, a metadata-free LLM-based automated feature engineering method that uses SHAP values and implicit search trajectories.

Xuan Zheng, Kento Uchida, and Shinichi Shirakawa: SIGMA: SHAP-Guided Implicit-Trajectory Generation for Metadata-Free LLM-Based AutoFE, The 18th Asian Conference on Machine Learning (ACML 2026), Melbourne, Australia, December 1-4, 2026. (Accepted) [arXiv] [Code]

Our paper has been published in Swarm and Evolutionary Computation. This paper proposes a framework for ensuring the invariance of multi-objective evolutionary algorithms to order-preserving transformations of objective functions.

Yutaro Yamada, Kento Uchida, and Shinichi Shirakawa: A framework for ensuring invariance of multi-objective evolutionary algorithms to order-preserving transformation of objective functions, Swarm and Evolutionary Computation, Vol. 108, 102527, Sep. 2026. [DOI]

Our paper has been published in Journal of Human Hypertension. This paper develops and evaluates prompt design methods for predicting atherosclerosis using large language models. This research is joint work with Prof. Yano (Juntendo University) and other collaborators.

Hibiki Murase, Kei Hiroshima, Kento Uchida, Mizuki Ohashi, Hiroyuki Daida, Naoki Kashihara, Naoki Nakashima, Anthony J. Viera, Kentaro Fujishiro, Hyeon Chang Kim, Shinichi Shirakawa, and Yuichiro Yano: Atherosclerosis prediction with large language models via prompt engineering, Journal of Human Hypertension, Sep. 2026. [DOI]

Our paper has been published in IEEE ACCESS. This paper proposes a surrogate benchmark for practical scenarios in the joint optimization of neural architecture and hyperparameters.

Yoichi Hirose, Kento Uchida, and Shinichi Shirakawa: NAS-HPO-Bench-II-Plus: A Surrogate Benchmark Dataset toward Practical Scenarios in Joint Neural Architecture Search and Hyperparameter Optimization, IEEE Access, Vol. 14, pp. 108050-108062, Jul. 2026. [DOI]

The collaborative paper with Prof. Akimoto’s group has been accepted to the 19th International Conference on Parallel Problem Solving From Nature (PPSN 2026). This paper proposes a bilevel optimization framework for mixed categorical-continuous black-box optimization.

• Marc Ong, Shinichi Shirakawa, Youhei Akimoto, Mixed-Categorical Black-Box Optimization via Information-Geometric Bilevel Decomposition, 19th International Conference on Parallel Problem Solving From Nature (PPSN 2026),Trento, Italy, August 29 - September 2, 2026. (Accepted) [arXiv]

Our paper has been accepted to the 28th International Conference on Pattern Recognition (ICPR 2026). This paper proposes a model-merging-based continual learning method that can consider performance preferences.

• Kei Hiroshima, Kento Uchida, Shinichi Shirakawa: Tunable MAGMAX: Preference-Aware Model Merging for Continual Learning, 28th International Conference on Pattern Recognition (ICPR 2026), Lyon, France, August 17-22, 2026.

Our papers (four full papers) have been accepted to the Genetic and Evolutionary Computation Conference (GECCO 2026) (San José, Costa Rica (hybrid), July 13-17, 2026).

• Haruhito Nakagawa, Kento Uchida, and Shinichi Shirakawa: Evaluation of Element-wise Effectiveness Estimation for Augmented Lagrangian CMA-ES (Accepted as a Full Paper)
• Sota Hamada, Yutaro Yamada, Kento Uchida, and Shinichi Shirakawa: Hierarchical Evolution Strategy for Optimization of Sharp Ridge (Accepted as a Full Paper)
• Kento Uchida, Ryoki Hamano, Masahiro Nomura, and Shinichi Shirakawa: Adaptive Stochastic Natural Gradient Method for Safe Optimization on Binary Space (Accepted as a Full Paper)
• Ryoki Hamano, Kento Uchida, and Shinichi Shirakawa: Convergence Analysis of Evolution Strategies for Mixed-Integer Optimization (Accepted as a Full Paper)

Yutaro Yamada (Doctoral Student) was selected as an Outstanding Student at Evostar 2026. Congratulations!