Kotaro Yoshida

Kotaro Yoshida

M.Sc. Student in Engineering

Institute of Science Tokyo

I am a master's student at the Institute of Science Tokyo, supervised by Prof. Konstantinos Slavakis. I am interested in interpreting large-scale models and leveraging that understanding to efficiently align, control, and edit them.

Previously, I focused on the theoretical understanding of pre-trained models and applied these insights to model merging. I have also worked on explaining generalization, out-of-distribution generalization, and confidence calibration through the lens of learning theory.

I was a research intern at A*STAR and ProPlace Inc. I spent a year at the National University of Singapore as an exchange student. Currently, I am a student researcher at Sakana AI.

Email GitHub Google Scholar X LinkedIn CV (Jan 2026)

News

Feb 2026 Joined Sakana AI as a student researcher, working on LLM Agents.
Jan 2026 Two first-authored papers accepted at ICLR 2026: "DisTaC: Conditioning Task Vectors via Distillation for Robust Model Merging" and "On Fairness of Task Arithmetic: The Role of Task Vectors".
Jun 2025 Started research internship at A*STAR (Singapore) under the SIPGA award.
Apr 2025 Paper accepted at TMLR 2025: "An Empirical Study of Pre-trained Model Selection for Out-of-Distribution Generalization and Calibration".
Apr 2025 Received the Thesis Proposal Excellence Award at Institute of Science Tokyo.
Jan 2025 First-authored paper accepted at ICLR 2025: "Mastering Task Arithmetic: τJp as a Key Indicator for Weight Disentanglement".
Apr 2025 Started M.Sc. at the Institute of Science Tokyo.
Mar 2025 Graduated B.Sc. from the Institute of Science Tokyo with the Outstanding Student Award.
Jun 2024 First-authored paper accepted at TMLR 2024: "Towards Understanding Variants of Invariant Risk Minimization from the Perspective of Calibration".

Publications

* denotes equal contribution.

International Conference

  • ICLR 2026

    "DisTaC: Conditioning Task Vectors via Distillation for Robust Model Merging"

    Kotaro Yoshida, Yuji Naraki, Takafumi Horie, Ryotaro Shimizu, Hiroki Naganuma

    The 14th International Conference on Learning Representations

    arXiv OpenReview GitHub
  • ICLR 2026

    "On Fairness of Task Arithmetic: The Role of Task Vectors"

    Laura Gomezjurado Gonzalez*, Kotaro Yoshida*. Hiroki Naganuma*, Takafumi Horie, Yuji Naraki, Ryotaro Shimizu

    The 14th International Conference on Learning Representations

    arXiv OpenReview GitHub
  • ICLR 2025

    "Mastering Task Arithmetic: τJp as a Key Indicator for Weight Disentanglement"

    Kotaro Yoshida, Yuji Naraki, Yoshikazu Ikeda, Takafumi Horie, Ryosuke Yamaki, Ryotaro Shimizu, Yuki Saito, Julian McAuley, Hiroki Naganuma

    The 13th International Conference on Learning Representations

    OpenReview GitHub

Journal

  • TMLR 2025

    "An Empirical Study of Pre-trained Model Selection for Out-of-Distribution Generalization and Calibration"

    Hiroki Naganuma, Ryuichiro Hataya, Kotaro Yoshida, Ioannis Mitliagkas

    Transactions on Machine Learning Research

    arXiv OpenReview GitHub
  • TMLR 2024

    "Towards Understanding Variants of Invariant Risk Minimization from the Perspective of Calibration"

    Kotaro Yoshida, Hiroki Naganuma

    Transactions on Machine Learning Research

    arXiv OpenReview GitHub

Preprint / Under Review

  • 2025

    "How Does Preconditioning Guide Feature Learning in Deep Neural Networks?"

    Kotaro Yoshida, Atsushi Nitanda

    Under Review

    arXiv GitHub
  • 2025

    "Robust Invariant Representation Learning by Distribution Extrapolation"

    Kotaro Yoshida, Konstantinos Slavakis

    arXiv Preprint

    arXiv
  • 2024

    "Augmenting NER Datasets with LLMs: Towards Automated and Refined Annotation"

    Yuji Naraki*, Ryosuke Yamaki*, Yoshikazu Ikeda, Takafumi Horie, Kotaro Yoshida, Ryotaro Shimizu, Hiroki Naganuma

    arXiv Preprint

    arXiv

Experience

A*STAR

Singapore

Singapore International Pre-Graduate Award (SIPGA) · Supervisor: Atsushi Nitanda

Jun 2025 – Sep 2025

Topic: Learning Theory and Model Merging

ProPlace Inc

Tokyo, Japan

Research Intern · Supervisor: Hiroki Naganuma

Jun 2023 – Jan 2026

Topic: Reliable and Efficient AI

A*STAR

Singapore

Research Intern · Supervisor: Atsushi Nitanda

Feb 2024 – May 2024

Research internship during exchange program at NUS

Service

Peer review.

  • 2025
    ICLR 2025 — Reviewer
  • 2026
    IEEE Transactions on Multimedia — Reviewer

Awards & Honors

Academic

  • 2025
    Thesis Proposal Excellence Award — Institute of Science Tokyo
  • 2025
    Outstanding Student Award — Institute of Science Tokyo (B.Sc.)
  • 2025
    Singapore International Pre-Graduate Award (SIPGA) — A*STAR
  • 2024
    Student Encouragement Award — IPSJ 86th National Convention

Competitions

  • 2024
    Bronze Medal (Top 4.5%) — SIGNATE SMBC Group GREEN×DATA Challenge
  • 2023
    Gold Medal (10th place, Top 1.6%) — SIGNATE Student Cup 2023
  • 2023
    Bronze Medal (Top 10%) — Kaggle ICR – Identifying Age-Related Conditions

Fellowships & Grants

  • 2021–Present
    HIOKI Scholarship — HIOKI E.E. Corporation
  • 2023–26
    TSUBAME Encouragement Program (3 times, 12,000 GPU hours total) — Tokyo Institute of Technology
  • 2023–24
    Support Scholarship for Study Abroad — Japan Student Services Organization (JASSO)

Education

Institute of Science Tokyo

Tokyo, Japan

M.Sc. in Engineering · Supervisor: Prof. Konstantinos Slavakis

Apr 2025 – Mar 2027 (Expected)

Topic: Efficient AI and Interpretability

Institute of Science Tokyo

Tokyo, Japan

B.Sc. in Engineering · Supervisor: Prof. Konstantinos Slavakis

Apr 2021 – Mar 2025

Topic: Out-of-Distribution Generalization and Calibration in Deep Learning

National University of Singapore

Singapore

Exchange Program in Computer Science

Aug 2023 – May 2024

Contact

Feel free to reach out if you are interested in my research or potential collaboration.

Email yoshida.k.0253[at]m.isct.ac.jp