Takuya Kataiwa
片岩 拓也Fourth-year student, Department of Information and Social Sciences,
Faculty of Informatics, Shizuoka University
Through research and implementation, I explore the relationship between people and computers. I want to create things no one has built before, and make existing large, complex systems faster, more scalable, and safer. My goal in both is to expand what people and computers can do.
Born in 2004, Ishikawa, Japan
What I'm working on
Inference optimization / Matter and computation
Research interests
Usable security / LLM security
Work & activities
Research Achievements
International Conferences
- Invisible Consent: Browser Automation by LLM Agents is Not Privacy-awareRyota Saito, Takuya Kataiwa, Minato Takashima, Tetsushi Ohki. Extended Abstracts of the ACM CHI Conference on Human Factors in Computing Systems (CHI EA '26). 2026.
- Secure and Usable Authentication Using Avatar Expression Blendshapes in Virtual RealityNatsuki Nagai, Tetsuro Takahashi, Takuya Kataiwa, Masakatsu Nishigaki, Tetsushi Ohki. Extended Abstracts of the ACM CHI Conference on Human Factors in Computing Systems (CHI EA '26). 2026.
Preprint
- Measuring Intrinsic Dimension of Token EmbeddingsTakuya Kataiwa, Cho Hakaze, Tetsushi Ohki.[arXiv] [Cited by 9]
Domestic Conferences (5)
- Invisible Consent: Browser Automation by LLM Agents is Not Privacy-awareRyota Saito, Takuya Kataiwa, Masakatsu Nishigaki, Tetsushi Ohki. Computer Security Symposium 2025. 2025.
- Measuring Intrinsic Dimension of Token EmbeddingsTakuya Kataiwa, Cho Hakaze, Tetsushi Ohki. The 39th Annual Conference of the Japanese Society for Artificial Intelligence. 2025.
- Simulator for Social Norm Formation by Aggregating Individual Behavioural DecisionsJun Tajima, Takuya Kataiwa, Kenya Sasaki, Junya Morita. The 39th Annual Conference of the Japanese Society for Artificial Intelligence. 2025.
- Measuring Intrinsic Dimension of Embedding RepresentationsTakuya Kataiwa, Cho Hakaze, Tetsushi Ohki. The 31st Annual Meeting of the Association for Natural Language Processing. 2025.
- A Method Using Dynamic Graph Neural Networks for Emotional Contagion in Local CommunitiesTakuya Kataiwa, Junya Morita. The 21st Informatics Workshop. 2023.
Poster Presentations (5)
- Effects of Expression Types in VR Facial Expression Recognition on Authentication Accuracy and UsabilityNatsuki Nagai, Tetsuro Takahashi, Takuya Kataiwa, Masakatsu Nishigaki, Tetsushi Ohki. The 2026 Symposium on Cryptography and Information Security. 2026.
- Evaluating Scalability in LLM-based Decompilation Using Source Code ObfuscationRyota Saito, Takuya Kataiwa, Makoto Yamaguchi, Yoshinobu Kano, Tetsushi Ohki. The 20th Young Researchers' Symposium on Natural Language Processing. 2025. Sponsor Award.
- Does Training Order Affect Transfer Performance and Internal Representations in Multilingual Continual Pretraining?Makoto Yamaguchi, Takuya Kataiwa, Yoshinobu Kano. The 20th Young Researchers' Symposium on Natural Language Processing. 2025.
- Hypothesis Exploration of Kunji-dominant Volumes of the Man'yōshū Using KL Divergence of n-gram ModelsShiori Shiratori, Takuya Kataiwa. The 20th Young Researchers' Symposium on Natural Language Processing. 2025.
- Enhancing LLM4Decompile with Obfuscated Source Code for Optimized Assembly (poster)Saito Ryota, Takuya Kataiwa, Tetsushi Ohki. The Network and Distributed System Security (NDSS) Symposium 2025.
Miscellaneous
- CPC Spring Camp 2025: A Student Participation ReportTakuya Kataiwa, Hana Kubota, Momoha Hirose. Journal of the Robotics Society of Japan, Vol.44, No.5, pp.488–491. 2026.
Selected Programs
Grants, Selected Projects, and Challenges
- MITOU IT Talent Scout and Development Program FY2026 — AcceptedTheme: Communication augmentation via tactile gel interfaces (PM: Oka)
- Kansai Tech Creator Challenge (K-TCC) 2025Theme: Holoprax: Extending cognition and the body to directly connect thought and action (PM: Hirao)
Fellowships and Research Development Programs
- Cybozu Labs Youth, 14th and 15th cohortsTheme: Developing a measurement framework for invisible consent by AI agents (Mentor: Nakatani)
- Nippon Foundation HUMAI Program, 1st and 2nd cohortsTheme: A theory of privacy centered on consent in the age of AI agentsDistinguished graduate (Excellence Award, 1st HUMAI Essay Contest)
- ADvance Lab, 2nd cohortTheme: Ensuring the safety and reliability of machine-learning model memory through geometry
- Security Camp National 2025, Class D
- 100 Program, 10th cohort (Summer 2026) — CompletedGrand Prize, Interview Excellence Award, and Hardware Excellence Award
Awards
University Awards
- Shizuoka University President's Award (FY2025)Mar 2026 | Shizuoka University
Hackathons
- OpenAI Codex Student Builder Fest 2026 — 1st PlaceSep 2026 | Hosted by OpenAI
- Open HackU 2024 Tokyo — Excellence AwardFeb 2024 | 6.7% (2/30) | Hosted by LY Corporation [Press Release]
Ideathons
- Urban Data Challenge 2025 — Gold AwardFeb 2026 | 0.4% (1/229)
- Government Reform Student Ideathon and Hackathon — Technology Utilisation AwardAug 2025 | 11.1% (1/9) | Hosted by the Cabinet Secretariat [Press Release]
Business Contests
- IVS SPIKATHON 2026 — WinnerJul 2026 | Hosted by 4S Inc. (IVS2026 hosted by the IVS KYOTO Executive Committee) [Press Release 1] [Press Release 2] [Press Release 3] [Press Release 4] [Press Release 5]
Other Awards
- Gikuhaku 2026 vol.2 — Sponsor Award (Works Human Intelligence Award)Sep 2026 | Hosted by Supporters
- Tongali Business Plan Contest 2026 — 3rd Place (Tongali Award), Team I-ArmJun 2026 | 18.8% (3/16) | Hosted by Tongali
- Gikuhaku 2025 vol.5 — Sponsor Award (WingArc 1st Award)Dec 2025 | 13.0% (3/23) | Hosted by Supporters
- Tekuno Ai 2025, University Student Division — Fighting Spirit AwardNov 2025 | Hosted by the Tekuno Ai Executive Committee
- Hamamatsu City Well-being Idea Contest, University/Graduate/Vocational Student Division — Excellence AwardNov 2025 | Hosted by LODU Inc. [Press Release] [News Article]
- Shizuoka University Ideathon and Datathon 2025 — Grand PrizeJun 2025 | 16.7% (1/6) | Co-hosted by Hamamatsu City and Shizuoka University
- Challenge Open Governance (COG 2024) — Semi-finalistFeb 2025 | 22.7% (25/110) | Hosted by the University of Tokyo
Open Source Contributions
Selected contributions (merged)
- PyTorch ☆ 103,177: Parallel Col2im, less memory [PR #198123]
- OpenCV ☆ 90,990: Reduce coordinate work and add SIMD in CPU transposed convolution [PR #30001]
- ONNX Runtime ☆ 21,943: Optimize Col2im, Im2col, Transpose, and 1D MaxPool [PR #32698] [PR #32800] [PR #32817] [PR #32891] [PR #32936]
Other merged contributions (8 PRs)
- ncnn ☆ 23,861: Reduce transposed-convolution memory access [PR #7015]
- MNN ☆ 16,156: Optimize transposed convolution [PR #4897] [PR #4912]
- Burn ☆ 15,966: Reuse buffers and parallelize transposed convolution [PR #5761]
- MLX Audio ☆ 7,991: Optimize Qwen3 speech-recognition timestamp repair [PR #982]
- tract ☆ 3,070: Optimize transposed-convolution memory access [PR #2909]
- ESP-DL ☆ 1,148: Optimize convolution and tensor operations [PR #332] [PR #334]
Proposed (9 PRs)
- vLLM ☆ 93,389: Reduce register spills in CPU AVX2 prefill attention [PR #58514]
- ONNX Runtime ☆ 21,943: Fuse CPU linear Resize and ArgMax [PR #33032]
- Candle ☆ 21,087: Optimize 1D transposed convolution [PR #3988]
- vLLM-Omni ☆ 7,079: Optimize timestamp decoding, VAE tile blending, and video buffer writes [PR #8230] [PR #8114] [PR #8113]
- tract ☆ 3,070: Avoid repeated dependency scans in evaluation order [PR #2957]
- ESP-DL ☆ 1,148: Weight reuse and argument placement [PR #333] [PR #335]
Management Experience
- 2026–: Japan Committee for the International Olympiad in Artificial Intelligence (JOAI Committee)
- 2024–: CPC Spring Camp Executive Committee / Symbol Emergence Outreach Team
History
- 2024–2025: ANCHOR (organized a reading group on neuroscience and cognitive science)
- 2025: ChoiAca (ran a site aggregating research event information)
- 2024–2026: IYNA (advisor)