Hiroaki Yamagiwa
My name is Hiroaki Yamagiwa. I'm a research scientist at SB Intuitions Corp.
Research Interests
large language models, data for language models, and mechanistic interpretability.Profiles
Google Scholar | ๐ | GitHub | CVContact
Work Experience
- Research Scientist SB Intuitions Corp. 1st Apr. 2026 – Present
- Assistant Professor Shimodaira Lab, Graduate School of Informatics, Kyoto University. 1st Apr. 2025 – 31st Mar. 2026
Education
- PhD in Informatics Graduate School of Informatics, Kyoto University 2022 – 2025 Under the supervision of Professor Hidetoshi Shimodaira.
- MSc in Informatics Graduate School of Informatics, Kyoto University 2020 – 2022 Under the supervision of Professor Hidetoshi Shimodaira.
- BSc in Mathematics Department of Mathematics, Kyoto University 2015 – 2020 Under the supervision of Program-Specific Senior Lecturer (at that time) Naoto Nakano.
Publications
(*) denotes equal contribution.
2026
- See Every Block Once: Variance-Reduced Post-Training for Masked Diffusion Language Models. In Proceedings of the 15th International Joint Conference on Natural Language Processing and the 5th Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (AACL-IJCNLP 2026), pp. 5836–5857, 2026.
- Measuring Affinity between Attention-Head Weight Subspaces via the Projection Kernel. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP 2026), pp. 34591–34660, 2026.
- Length and Meaning for Language Model Maps. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP 2026), pp. 34472–34487, 2026.
- Domain Mixture Design via Log-Likelihood Differences for Aligning Language Models with a Target Model. In Findings of the Association for Computational Linguistics: EMNLP 2026, pp. 30483–30499, 2026.
- Establishing a Scale for Kullback–Leibler Divergence in Language Models Across Various Settings. In Findings of the Association for Computational Linguistics: ACL 2026, pp. 23223–23248, 2026.
2025
- Likelihood Variance as Text Importance for Resampling Texts to Map Language Models. In Findings of the Association for Computational Linguistics: EMNLP 2025, pp. 9453–9465, 2025.
- Mapping 1,000+ Language Models via the Log-Likelihood Vector. 2025. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), pp. 32983–33038, 2025. ๐Outstanding Paper Award (26/8360)
- Quantifying Lexical Semantic Shift via Unbalanced Optimal Transport. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), pp. 15913–15933, 2025.
- Predicting drug-gene relations via analogy tasks with word embeddings. Scientific Reports 15, 17240, 2025.
- Norm of Mean Contextualized Embeddings Determines their Variance. In Proceedings of the 31st International Conference on Computational Linguistics (COLING 2025), pp. 7778–7808, 2025.
- Revisiting Cosine Similarity via Normalized ICA-transformed Embeddings. In Proceedings of the 31st International Conference on Computational Linguistics (COLING 2025), pp. 7423–7452, 2025.
2024
- Understanding Higher-Order Correlations Among Semantic Components in Embeddings. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 2883–2899, 2024.
- Axis Tour: Word Tour Determines the Order of Axes in ICA-transformed Embeddings. In Findings of the Association for Computational Linguistics: EMNLP 2024, pp. 477–506, 2024.
- Shimo Lab at "Discharge Me!": Discharge Summarization by Prompt-Driven Concatenation of Electronic Health Record Sections. In Proceedings of the 23rd Workshop on Biomedical Natural Language Processing, pp. 645–657, 2024.
- Zero-Shot Edge Detection with SCESAME: Spectral Clustering-based Ensemble for Segment Anything Model Estimation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV 2024) Workshops, pp. 541–551, 2024.
2023
- Discovering Universal Geometry in Embeddings with ICA. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 4647–4675, 2023.
- Improving word mover's distance by leveraging self-attention matrix. In Findings of the Association for Computational Linguistics: EMNLP 2023, pp. 11160–11183, 2023.
Preprints
-
J-tau: A Fully Localized Japanese Benchmark for Tool-Agent-User Interaction.
arXiv 2026.
-
Industry Contributions
- Sarashina3.2 mini thinking 2026 Project member. A Japanese LLM with thinking capability, developed at SB Intuitions. ๅฝ็ฃLLM Sarashina3.2 mini thinking ใฎ็ดนไป (Introducing Sarashina3.2 mini thinking, a Japanese LLM), SB Intuitions TECH BLOG
Research Activities
Reviewer
2026
2025
- Mechanistic Interpretability Workshop at NeurIPS 2025
- ACL Rolling Review
- ACL 2025 Student Research Workshop
- Natural Language Processing (renamed from "Natural Language Engineering" in 2024)
- The 31st International Conference on Computational Linguistics (COLING 2025)
2024
Activities
Kaggle
I became a Kaggle Competitions Expert in 2021. My Kaggle profile is here.