Tatsuro Inaba is a first-year Ph.D. student in the NLP Department at
MBZUAI
, supervised by Prof. Kentaro Inui and Prof. Gus Xia.
His research interests include training dynamics and hierarchical capabilities of both language and music models.
- Email: tatsuro.pianooo[at]gmail.com
- X (Twitter)
- Google Scholar
- GitHub
News
- 09/2025: My first-author paper was accepted to AI4Music workshop co-located with NeurIPS 2025.
- 09/2025: Our paper was accepted to NeurIPS 2025.
- 08/2025: Joined MBZUAI as a Ph.D. student.
- 08/2025: My first-author paper was accepted to EMNLP 2025.
- 01/2025: Our paper was accepted to NAACL 2025.
Career
Education
- 08/2025-present: Ph.D. student, NLP Department, Mohamed Bin Zayed University of Artificial Intelligence (MBZUAI).
- 04/2025-08/2025: Doctoral program, Graduate School of Information Sciences, Tohoku University (withdrew Aug 2025).
- 04/2023-03/2025: M.S. student, Graduate School of Informatics, Kyoto University.
- 04/2019-03/2023: B.Eng. student, Undergraduate School of Electrical and Electronic Engineering, Kyoto University.
Professional Experience
- 12/2024-08/2025: Research Assistant, Research and Development Center for Large Language Models, National Institute of Informatics.
- 04/2025-08/2025: JSPS Research Fellow (DC1), Tohoku University.
- 01/2024-02/2024: Visiting Student, MBZUAI. [blog]
- 11/2023-11/2024: Research Assistant, Kyoto University.
- 09/2023-10/2023: Machine Learning Engineer (Intern), Recruit Co., Ltd.
- 08/2023-09/2023: Research Intern, Preferred Networks, Inc. [blog]
- 11/2022-07/2023: Research Engineer (Part-time), DATAGRID Inc.
Grants/Awards
- 04/2025-08/2025: JSPS Research Fellowship (DC1) (708/4959 = 14.3%)
- 08/2024: Student Encouragement Award, 141st IPSJ SIGMUS Conference in Japan (SIGMUS 141).
- 08/2023: Encouragement Award (20/139 = 14.4%), The 18th Symposium of Young Researcher Association for NLP Studies in Japan (YANS 2023).
- 03/2023: Outstanding Award (13/579 = 2.2%), 29th Annual Meeting of the Language Processing Society of Japan (NLP 2023).
Publications
Refereed Papers
-
Tatsuro Inaba and Kentaro Inui.
"Why Do Music Models Plagiarize? A Motif-Centric Perspective"
NeurIPS 2025 Workshop on AI for Music
-
Mengyu Ye, Jun Suzuki, Tatsuro Inaba, and Tatsuki Kuribayashi.
"Transformer Key-Value Memories Are Nearly as Interpretable as Sparse Autoencoders."
The Thirty-ninth Annual Conference on Neural Information Processing Systems (NeurIPS 2025)
[arXiv, project page] -
Tatsuro Inaba, Go Kamoda, Kentaro Inui, Masaru Isonuma, Yusuke Miyao, Yohei Oseki, Yu Takagi, and Benjamin Heinzerling.
"How a Bilingual LM Becomes Bilingual: Tracing Internal Representations with Sparse Autoencoders."
The 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP 2025 Findings)
[arXiv] -
Go Kamoda, Benjamin Heinzerling, Tatsuro Inaba, Keito Kudo, Keisuke Sakaguchi, and Kentaro Inui.
"Weight-based Analysis of Detokenization in Language Models: Understanding the First Stage of Inference Without Inference."
The 2025 Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL 2025 Findings)
[arXiv] -
Tatsuro Inaba, Kazuyoshi Yoshii, and Eita Nakamura.
"On the Importance of Time and Pitch Relativity for Transformer-Based Symbolic Music Generation."
The 16th Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA 2024)
[paper] -
Tatsuro Inaba, Hirokazu Kiyomaru, Fei Cheng, and Sadao Kurohashi.
"MultiTool-CoT: GPT-3 Can Use Multiple External Tools with Chain of Thought Prompting"
The 61st Annual Meeting of the Association for Computational Linguistics (ACL 2023 Main)
[code, arXiv, poster]
Preprints
-
Ryosuke Takahashi, Tatsuro Inaba, Kentaro Inui, and Benjamin Heinzerling.
"TopK Language Models.", Jun 2025.
[arXiv]
Domestic Conference (Japan)
-
神林 励旺, 稲葉 達郎, Benjamin Heinzerling, 坂口 慶祐, 磯沼 大.
「マルチモーダルLLMのモダリティ間共有表現の解明」
NLP若手の会 (YANS) 第20回シンポジウム, 2025. 奨励賞
-
稲葉 達郎, 乾 健太郎, 宮尾 祐介, 大関 洋平, Benjamin Heinzerling, 高木 優.
「スパースオートエンコーダーを用いた大規模言語モデルのチェックポイント横断分析」
言語処理学会第31回年次大会, 2025.
[paper] -
鴨田 豪, Benjamin Heinzerling, 稲葉 達郎, 工藤 慧音, 坂口 慶祐, 乾 健太郎.
「言語モデルのパラメータから探るDetokenizationメカニズム」
言語処理学会第31回年次大会, 2025.
[paper] -
稲葉 達郎, 吉井 和佳, 中村 栄太.
「音楽生成における時間と音高相対性の重要性」
第141回音楽情報科学研究発表会 (SIGMUS), 2024. 学生奨励賞
[paper] -
稲葉 達郎, 藤井 拓郎, 小原 涼馬, 柴田 幸輝.
「三言語モデル寄れば文殊の知恵を」
NLP若手の会 (YANS) 第18回シンポジウム, 2023. 奨励賞
[poster] -
稲葉 達郎
「大規模言語モデルに基づく複数の外部ツールを利用した推論フレームワーク」の研究
自然言語処理 30巻 3号 p.1100-1104, 2023.
[paper] -
稲葉 達郎, 清丸 寛一, Fei Cheng, 黒橋 貞夫.
「大規模言語モデルに基づく複数の外部ツールを利用した推論フレームワーク」
言語処理学会第29回年次大会, 2023. 優秀賞
[paper, poster]
Talks
- 06/2025: Talk and tutorials (PyTorch/Transformer/Music Generation) at Kyushu University. [talk slides]
- 03/2025: Talk at Tohoku NLP Group. [slides]
Activities
- 09/2025: 研究紹介, 第24回LLM勉強会, "How a Bilingual LM Becomes Bilingual: Tracing Internal Representations with Sparse Autoencoders", [slides]
- 03/2025: 研究紹介, 第17回LLM勉強会, "スパースオートエンコーダを用いた大規模言語モデルのチェックポイント横断分析", [slides]
- 08/2024: 論文紹介, 第16回最先端NLP勉強会, "Language-Specific Neurons: The Key to Multilingual Capabilities in Large Language Models", [slides]
- 09/2023: 学会記事, 「大規模言語モデルに基づく複数の外部ツールを利用した推論フレームワーク」の研究.自然言語処理 30巻 3号 p.1100-1104, [paper]
- 08/2023: 論文紹介, 第15回最先端NLP勉強会, "Backpack Language Models", [slides]