Skip to main content
Yusuke Kimura
Menu
Language: EN
Portrait of Yusuke Kimura

Yusuke Kimura

Kimura Yusuke

NLP · LLM Post-training · Cross-tokenizer Distillation

I study post-training methods for adapting large language models efficiently and reliably to different tasks and deployment settings. My current focus is knowledge transfer between models with different vocabularies and tokenizers, including learning under limited compute.

See all research outputs View CV

Highlights

Activity timeline

67 entries2017–2026

Publications, presentations, awards, and research support, from newest to oldest.

Scroll within the panel to browse the complete history.

See all research outputs

202617 entries

  1. Teaching

    I have started teaching “Practical Exercises for Data Science and AI–7: Fundamentals of Natural Language Processing” at Doshisha University.(opens in a new tab)

    Fall 2026 · 2 credits · A general education course open to first-year students and above, covering NLP and LLM fundamentals, classification, retrieval, generation, and fine-tuning. On-demand lectures with an in-person final test.

  2. Talk

    I presented “Find Your Optimal Teacher: Personalized Data Synthesis via Router-Guided Multi-Teacher Distillation” at the ACL 2026 paper reading meeting at Nagoya University.(opens in a new tab)

    A paper discussion on selecting a suitable teacher model for each query in knowledge distillation · Nagoya University

  3. Participation

    I attended the Tokai–Kansai Database Workshop 2026 (DBWS 2026) as a faculty member, listening to the presentations.(opens in a new tab)

    A joint workshop where university laboratories in data engineering from the Tokai and Kansai regions gather for poster presentations and research exchange · Shizuoka University, Hamamatsu Campus

  4. Update

    I am serving on the programme committees of SoICT 2026 and iiWAS 2026.

    Both conferences take place in December 2026.

  5. Paper

    The research output “ByteTop-k OPD: バイト列表現に基づく語彙の異なる LLM 間のオンポリシ蒸留” was published.

    IPSJ SIG-NL Technical Report 2026-NL-269

  6. Media

    The official YANS 2026 report has been published, including the selection rationale for my ELYZA Award.(opens in a new tab)

    “On-Policy Distillation between LLMs with Different Vocabularies” · Recognized for directly addressing vocabulary differences, practical potential, and prospects for further development.

  7. Personal award

    I presented the poster “On-Policy Distillation between LLMs with Different Vocabularies” and received the ELYZA Award.(opens in a new tab)

    The 21st YANS Symposium (YANS 2026) · Sendai International Center · Link opens the official report with the award rationale

  8. Coauthor award

    平子 翔太 received the 2026 IPSJ Yamashita SIG Research Award.(opens in a new tab)

    Related publication: LoRAマルチタスク学習における更新空間の分離による学習偏りの抑制

    IPSJ SIG-IFAT Technical Report 2026-IFAT-162(4)

  9. Paper

    The research output “LLM を用いた Web アプリケーションのチューニング” was published.

    2026 Tokai-Section Joint Conference on Electrical, Electronics, Information, and Related Engineering (2B-1-4)

  10. Paper

    The research output “機械学習ベースの索引チューニングにおける説明可能 AI の適用” was published.

    2026 Tokai-Section Joint Conference on Electrical, Electronics, Information, and Related Engineering (2B-1-3)

  11. KAKENHI

    I received a JSPS KAKENHI Grant-in-Aid for Research Activity Start-up.(opens in a new tab)

    Project 26K25590, "Controlling the Allocation of Update Resources during Training for Lightweight Language-Model Adaptation to Under-Trained Domains" (Principal Investigator, FY2026–2027)

  12. Compute & shared use

    The project “Interference Analysis of LLM Adapter Learning under Composite Instructions and Continual Adaptation, and Construction of a High-Efficiency Experimental Platform” was selected for shared computing support.

    JHPCN Seed-Type Collaborative Research

  13. Compute & shared use

    The project “Interference Analysis of LLM Adapter Learning under Composite Instructions and Continual Adaptation, and Construction of a High-Efficiency Experimental Platform” was selected for shared computing support.

    Kyoto University Supercomputer Collaborative Research Program (Young/Female Researcher Track)

  14. Compute & shared use

    The project “Efficient Adapter Learning via Layer-wise Dynamic Quantization Driven by Training Progress” was selected for shared computing support.

    University of Tokyo Information Technology Center, Young/Female User Recommendation

  15. Paper

    The research output “LoRAマルチタスク学習における更新空間の分離による学習偏りの抑制” was published.

    IPSJ SIG-IFAT Technical Report 2026-IFAT-162(4)

  16. Coauthor award

    前川 知歩(opens in a new tab) received the Young Researcher Award (IPSJ SIG-NL).(opens in a new tab)

    Related publication: 社会的属性ごとに異なる内部表現を考慮した活性化ステアリングによるLLMの社会的バイアス緩和

    IPSJ SIG-NL Technical Report 2026-NL-267(31)

  17. Paper

    The research output “社会的属性ごとに異なる内部表現を考慮した活性化ステアリングによるLLMの社会的バイアス緩和” was published.

    IPSJ SIG-NL Technical Report 2026-NL-267(31)

202512 entries

  1. Personal award

    I received the “Money Forward Award(opens in a new tab)”.

    Money Forward, Inc.

  2. Personal award

    I received the “Director-General of Kinki Bureau of Telecommunications Award(opens in a new tab)”.

    Kinki Bureau of Telecommunications

  3. Personal award

    I received the “Outstanding Research Award(opens in a new tab)”.

    SIG Web Intelligence and Interaction

  4. Paper

    The research output “外部知識を優先した出力生成のための類似度カーネル平滑化手法” was published.

    21st Workshop on Web Intelligence and Interaction (WI2)

  5. Presentation

    I participated in the Tokai–Kansai Database Workshop 2025 (DBWS 2025) as a student and presented the poster “A Study on a Quantization Scheduler for Effective Adapter Training”.(opens in a new tab)

    Joint workshop involving multiple universities and laboratories in the Kansai and Tokai regions · Doshisha University, Imadegawa Campus

  6. Compute & shared use

    The project “Development of Efficient LLM Adapter Learning with a Quantization Scheduler” was selected for shared computing support.

    JHPCN Seed-Type Collaborative Research

  7. Compute & shared use

    The project “Development of Efficient LLM Adapter Learning with a Quantization Scheduler” was selected for shared computing support.

    Science Tokyo TSUBAME Young/Female User Support Program

  8. Paper

    The research output “検索対象文書集合を把握するためのコンテキスト内学習に基づいたクエリ書換え手法” was published.

    IEICE General Conference 2025, Junior & Student Poster Session

  9. Paper

    The research output “フリマアプリにおける購入時考慮属性の生成手法の提案” was published.

    Proceedings of DEIM 2025

  10. Paper

    The research output “食に関する異種グラフを用いた料理の栄養価推定” was published.

    Proceedings of DEIM 2025

  11. Paper

    The research output “開発コスト軽減化のためのマルチタスク強化学習を用いた系列推薦システム” was published.

    IPSJ Transactions on Databases (TOD)

  12. Paper

    The research output “RICGraph: A Recipe-Ingredient-Compound Graph for Estimating the Nutritional Value of Recipes” was published.

    41st IEEE International Conference on Data Engineering Workshops (ICDEW)

20248 entries

  1. Paper

    The research output “L3Masking: Multi-task Fine-tuning for Language Models by Leveraging Lessons Learned from Vanilla Models” was published.

    Proceedings of the 1st Workshop on Customizable NLP (CustomNLP4U) @ EMNLP

  2. Paper

    The research output “A Soft Actor-Critic Algorithm for Sequential Recommendation” was published.

    Lecture Notes in Computer Science

  3. Coauthor award

    洪 惠珍(opens in a new tab) received the Student Encouragement Award, 86th IPSJ National Convention.(opens in a new tab)

    Related publication: Soft Actor-Critic 強化学習を用いた系列推薦フレームワーク

    Proceedings of the 86th IPSJ National Convention

  4. Paper

    The research output “ドメイン適応のためのトークン単位の擬似尤度に基づくマスク戦略” was published.

    IPSJ SIG Technical Report

  5. Paper

    The research output “追加事前学習による日本語言語モデルの倫理性獲得の試み” was published.

    Proceedings of the 86th IPSJ National Convention

  6. Paper

    The research output “Soft Actor-Critic 強化学習を用いた系列推薦フレームワーク” was published.

    Proceedings of the 86th IPSJ National Convention

  7. Paper

    The research output “部分構造を考慮した化合物分散表現の食材分類タスクにおける効果” was published.

    DBSJ Japanese Journal

  8. Paper

    The research output “Construction of Ingredient Embedding Considering Both Cooking Recipes and Their Ingredients” was published.

    2024 IEEE International Conference on Big Data and Smart Computing (BigComp)

20236 entries

  1. Paper

    The research output “An Automatic Labeling Method for Subword-Phrase Recognition in Effective Text Classification” was published.

    Informatica

  2. Doctoral support

    I was selected for doctoral research support through “Doshisha University SPRING Program(opens in a new tab)”.

    Doshisha University Graduate School (competitive internal selection)

  3. Paper

    The research output “商品説明に着目したレビュー文からの教師なしキーフレーズ抽出” was published.

    Proceedings of DEIM 2023

  4. Paper

    The research output “文書分類における効果的なマルチタスク学習のための汎用的サブタスクの要件分析” was published.

    IEICE Technical Report

  5. Paper

    The research output “言語モデルに基づく知識グラフの構造を学習するリンク予測手法” was published.

    IPSJ SIG Technical Report

  6. Paper

    The research output “Ingredient Embeddings Constructed by Biased Random Walk on Ingredient-Compound Graph” was published.

    Procedia Computer Science

20229 entries

  1. Paper

    The research output “Multi-task Learning-based Text Classification with Subword-Phrase Extraction” was published.

    The 11th International Symposium on Information and Communication Technology (SoICT)

  2. Presentation

    I attended the Tokai–Kansai Database Workshop 2022 (DBWS 2022) as a student and presented the poster “A Multi-task Learning Framework for Document Classification Using Self-supervised Learning”.(opens in a new tab)

    Joint workshop of university laboratories in the Tokai and Kansai regions · Konan University, Hirao Memorial Seminar House

  3. Coauthor award

    吉丸 直希(opens in a new tab) received the Student Encouragement Award (IPSJ SIG-DBS).(opens in a new tab)

    Related publication: 部分構造を考慮した化合物分散表現の食材分類タスクにおける効果

    IPSJ SIG-IFAT Technical Report 2022-IFAT-148(31)

  4. Paper

    The research output “部分構造を考慮した化合物分散表現の食材分類タスクにおける効果” was published.

    IPSJ SIG-IFAT Technical Report 2022-IFAT-148(31)

  5. Doctoral support

    I was selected for doctoral research support through “Doshisha University Doctoral Fellowship for Young Researchers(opens in a new tab)”.

    Doshisha University Graduate School (competitive internal selection)

  6. Coauthor award

    吉丸 直希(opens in a new tab) received the Student Encouragement Award, 84th IPSJ National Convention.(opens in a new tab)

    Related publication: レシピでの共起確率を考慮した食材分散表現の構築

    Proceedings of the 84th IPSJ National Convention

  7. Paper

    The research output “前後文のアスペクトを考慮したレビュー文に対する教師ありアスペクト推定の高精度化” was published.

    IEICE General Conference 2022, Junior & Student Poster Session

  8. Paper

    The research output “レシピでの共起確率を考慮した食材分散表現の構築” was published.

    Proceedings of the 84th IPSJ National Convention

  9. Paper

    The research output “ストップフレーズ抽出を併用した文書分類” was published.

    Proceedings of DEIM 2022

20215 entries

  1. Presentation

    I attended the Tokai–Kansai Database Workshop 2021 (DBWS 2021) as a student and presented the poster “A Tokenizer for High-accuracy Information Retrieval”.(opens in a new tab)

    Joint workshop of university laboratories in the Tokai and Kansai regions · Online

  2. Coauthor award

    小森 雄太 received the Student Encouragement Award, 83rd IPSJ National Convention.(opens in a new tab)

    Related publication: 単語の概念関係を考慮したレビューの評価観点自動抽出手法

    83rd IPSJ National Convention

  3. Paper

    The research output “単語の概念関係を考慮したレビューの評価観点自動抽出手法” was published.

    83rd IPSJ National Convention

  4. Paper

    The research output “対話における意図分類を目的とした発話の分散表現構築法” was published.

    Proceedings of IEICE General Conference 2021

  5. Paper

    The research output “Automatic Terminology Extraction Using a Dependency-Graph in NLP” was published.

    Proceedings of the 11th International Conference on Innovations in Bio-Inspired Computing and Applications

20206 entries

  1. Paper

    The research output “単語埋め込みと名詞句の共起グラフを用いた教師なしキーフレーズ抽出手法の提案” was published.

    IPSJ SIG Technical Report (IFAT-140)

  2. Coauthor award

    笹田 大翔(opens in a new tab) received the Student Encouragement Award, 82nd IPSJ National Convention.(opens in a new tab)

    Related publication: データの複雑性を考慮した不均衡データに対するリサンプリング手法

    82nd IPSJ National Convention

  3. Paper

    The research output “データの複雑性を考慮した不均衡データに対するリサンプリング手法” was published.

    82nd IPSJ National Convention

  4. Paper

    The research output “物体検出技術を用いた古典籍画像からのテキスト領域切り出し手法” was published.

    IEICE General Conference 2020, Student Poster Session

  5. Paper

    The research output “係り受け構解析を用いたドメイン固有のキーワード抽出手法の提案” was published.

    IEICE General Conference 2020, Student Poster Session

  6. Paper

    The research output “A Resampling Method for Imbalanced Datasets Considering Noise and Overlap” was published.

    Procedia Computer Science

20193 entries

  1. Presentation

    I attended the Tokai–Kansai Database Workshop 2019 (DBWS 2019) as a student and presented the poster “Automatic Term Extraction Using Dependency and Anaphoric Relations”.(opens in a new tab)

    Joint workshop of university laboratories in the Tokai and Kansai regions · University of Hyogo, Kobe Campus for Information Science

  2. Personal award

    I received the “LegalForce Corporate Award”.

    LegalForce, Inc.

  3. Paper

    The research output “係り受け構造に着目した専門用語自動抽出手法の提案” was published.

    IEICE General Conference 2019, Student Poster Session

20171 entries

  1. Personal award

    I received the “Excellence Award”.

    Joint Council of Management Science Research Associations

Research

Primary focus

Knowledge Distillation between LLMs with Different Vocabularies and Tokenizers

I study post-training methods that align otherwise incomparable teacher and student outputs through a shared representation, enabling knowledge transfer between LLMs whose vocabularies and token boundaries differ.

Collaborative

Efficient Adapter Learning under Limited Compute

Adapting LLMs under tight compute budgets via quantization scheduling and training-progress-driven layer-wise dynamic quantization (collaborative, backed by competitive HPC allocations).

Collaborative

Reliable Adaptation: Bias Mitigation & Grounded Generation

Trustworthy LLM adaptation through social-bias mitigation with activation steering and knowledge-prioritized generation (collaborative).

Survey & research resources

Cross-Tokenizer OPD

A research catalogue of on-policy distillation between models with different tokenizers. Explore paper claims, methods, teacher–student pairs, official code, and citations.

Open research catalogue
Keywords
LLM Post-training · On-policy Distillation · Cross-tokenizer Knowledge Distillation · Efficient Adapter Learning · Natural Language Processing · Multi-task Learning · Text Classification · Keyphrase Extraction
Fields
Intelligent Informatics · Databases · Web & Service Informatics

Doctoral Thesis

A Study on Designing Annotation-Free Tasks for Domain Adaptation of Language Models in Text Classification

Yusuke Kimura

Graduate School of Culture and Information Science, Doshisha University (academic year 2025)

Thesis record (laboratory website) (opens in a new tab)

Awards

6 total

  1. ELYZA Award

    Young Researcher Association for NLP Studies (YANS)

    On-Policy Distillation between LLMs with Different Vocabularies

    Awarded for the poster presentation at the 21st YANS Symposium (YANS 2026).

    Official report and award rationale (opens in a new tab)

    Recipients: 木村 優介

  2. Money Forward Award

    Money Forward, Inc.

    AI-Collaborative Developmental Support System to Drive DX in Developmental Care Services

    CognitiBase business plan pitched in the Student Division of Miraino Pitch 2025, proposing how AI can be introduced into developmental care

    CognitiBase venture site (opens in a new tab)

    Recipients: 木村優介, 楠和馬

  3. Director-General of Kinki Bureau of Telecommunications Award

    Kinki Bureau of Telecommunications

    AI-Collaborative Developmental Support System to Drive DX in Developmental Care Services

    CognitiBase business plan pitched in the Student Division of Miraino Pitch 2025, proposing how AI can be introduced into developmental care

    CognitiBase venture site (opens in a new tab)

    Recipients: 木村優介, 楠和馬

  4. Outstanding Research Award

    SIG Web Intelligence and Interaction

    A Similarity-Kernel Smoothing Method for Knowledge-Prioritized Generation

    21st WI2 meeting report (opens in a new tab)

    Recipients: 森岡 葵子, 吉丸 直希, 木村 優介, 波多野 賢治

  5. LegalForce Corporate Award

    LegalForce, Inc.

    Automatic Term Extraction Using Dependency Structure

    Recipients: 木村 優介

  6. Excellence Award

    Joint Council of Management Science Research Associations

    Predictive Model for Customer Volume in Booth-Rental Hair Salons

    Recipients: 井上 由佳, 越前谷 勇典, 木村 優介, 窪田 茜, 山田 祐也

Research support

JSPS KAKENHI (PI) (1)

  1. Controlling the Allocation of Update Resources during Training for Lightweight Language-Model Adaptation to Under-Trained Domains (opens in a new tab)

    JSPS KAKENHI, Grant-in-Aid for Research Activity Start-up

Compute & shared-use programs (5)

  1. Interference Analysis of LLM Adapter Learning under Composite Instructions and Continual Adaptation, and Construction of a High-Efficiency Experimental Platform

    JHPCN Seed-Type Collaborative Research

  2. Interference Analysis of LLM Adapter Learning under Composite Instructions and Continual Adaptation, and Construction of a High-Efficiency Experimental Platform

    Kyoto University Supercomputer Collaborative Research Program (Young/Female Researcher Track)

  3. Efficient Adapter Learning via Layer-wise Dynamic Quantization Driven by Training Progress

    University of Tokyo Information Technology Center, Young/Female User Recommendation

    • Principal Investigator

  4. Development of Efficient LLM Adapter Learning with a Quantization Scheduler

    JHPCN Seed-Type Collaborative Research

    • Project no. EX25407

  5. Development of Efficient LLM Adapter Learning with a Quantization Scheduler

    Science Tokyo TSUBAME Young/Female User Support Program

Doctoral research support (internal selection) (2)

  1. Doshisha University SPRING Program (opens in a new tab)

    Doshisha University Graduate School (competitive internal selection)

    • SPRING-supported doctoral student

  2. Doshisha University Doctoral Fellowship for Young Researchers (opens in a new tab)

    Doshisha University Graduate School (competitive internal selection)

    • Fellowship recipient

Career / Education

Career

  1. Present

    Researcher, JST CREST (Strategic Basic Research Programs)

    Graduate School of Informatics, Nagoya University

  2. Teaching Assistant (Database Systems, Programming I–III)

    Graduate School of Culture and Information Science, Doshisha University

  3. Research Internship

    NTT Communication Science Laboratories

  4. Specialized Staff

    Faculty of Culture and Information Science, Doshisha University

Education

  1. Ph.D. Program, Culture and Information Science

    Doshisha University Graduate School

  2. M.S. Program, Culture and Information Science

    Doshisha University Graduate School

  3. B.A., Culture and Information Science

    Doshisha University

  4. General Course

    Hyogo Prefectural Takarazuka-Kita High School

Teaching

  1. Fall 2026

    Practical Exercises for Data Science and AI–7: Fundamentals of Natural Language Processing

    Doshisha University

    General education · First-year students and above · 2 credits · On-demand lectures with an in-person final test

    Syllabus (opens in a new tab)

Academic activities

Academic service

  1. SoICT2026 Program Committee

    Peer Review

  2. iiWAS2026 Program Committee

    Peer Review

  3. Present

    IPSJ-TOD Reviewer

    Peer Review

  4. IPSJ Journal Reviewer

    Peer Review

  5. EMNLP Workshop Reviewer

    Peer Review

  6. Wikipedia + Extended Named Entity Structuring Project (Task 1: Classification)

    Other RIKEN AIP, Language Information Access Technology Team

Memberships

  1. Present

    Association for Computational Linguistics (ACL)

Contact

kimura.yusuke.p2@f.mail.nagoya-u.ac.jp

Nagoya University