I am a researcher at the Sato Laboratory for Biomedical Data Science, School of Life Science and Technology, Institute of Science Tokyo. I am also a guest researcher at the Frith Lab, The University of Tokyo, and the founder of AlphaScience Lab, an initiative focused on expertise-guided, auditable, and responsible AI-assisted scientific discovery.
My current research focuses on AI-automated scientific workflows, AI research and evaluation, bioinformatics, and AI for Science. I am interested in building computational systems that help automate, audit, and accelerate scientific discovery.
Recent research includes audit-constrained protocols for targeted tests of LLM reasoning, minimal models separating shortcut-rule transition from cross-family OOD failure, controlled counterexamples to strong proxy-based explanations of OOD performance, and FastUMAP, a scalable dimensionality-reduction method based on bipartite landmark sampling. I also work on ID3 as a task-general input-data differentiable framework. While biomolecular sequence design is one current application, the core idea is meant to extend to general task settings where inputs, objectives, and evaluation signals can be made differentiable and auditable.
