Issa Sugiura
I am a first year PhD student at Institute of Science Tokyo, majoring in Computer Science.
I like to learn by building things.
Research
I work on vision-language models (VLMs) and evaluations (evals).
- VLMs: developing VLMs through the whole process, from building large-scale training data (WAON, Jagle) to training models (LLM-jp-4-VL 9B) and evaluating them (JAMMEval, HakushoBench, simple-evals-mm).
- Evals: building benchmarks for tasks that frontier models are expected to take on, such as financial analysis (EDINET-Bench, JFinBench) and running a business honestly (CoffeeBench).
Highlights
- First-author papers at ICLR, COLM, and NeurIPS (E&D) 2026, plus two at AACL 2026 Findings.
- Best Paper Award at NLP2026 (3/789, top 0.4%) and four other awards.
- Open datasets and models with 85K+ Hugging Face downloads; WAON, Jagle, and WAON-Bench are used by Noetra to build its multimodal foundation model.
Education
- Apr 2026 – present
Doctoral Program (Expected)
Institute of Science Tokyo, Japan
Course of Artificial Intelligence, Department of Computer Science, School of Computing
- Apr 2024 – Mar 2026
Master's Degree
Kyoto University, Japan
Course of Communications and Computer Engineering, Graduate School of Informatics
- Apr 2020 – Mar 2024
Bachelor's Degree
Osaka University, Japan · GPA 3.57/4.00
Software Science Course, Department of Information and Computer Sciences, School of Engineering Science
Experience
- Apr 2026 – present
Research Assistant · Institute of Science Tokyo, Japan
Research on VLMs.
- Dec 2024 – present
Student Intern · Sakana AI, Japan
Building benchmarks for tasks that frontier models are expected to take on.
- Led EDINET-Bench (ICLR 2026), a benchmark of fraud detection, earnings forecasting, and industry classification built from 10 years of Japanese annual reports; frontier LLMs barely beat logistic regression.
- Led CoffeeBench (NeurIPS 2026 E&D, with KPMG AZSA), where six LLM coding agents run a coffee supply chain for 90 simulated days. Under a revenue target, GPT-6 Astra proposes circular trades that inflate revenue, and whether they happen depends on the counterparty (Claude Sonnet 5.5 refuses).
- Led JFinBench, 321 questions from the 2026 editions of six Japanese certification exams in accounting, tax, financial planning, and banking, built with a human-in-the-loop coding-agent pipeline so it can be refreshed with each new exam.
- Feb 2024 – present
Research Assistant · National Institute of Informatics, Japan
Developing Japanese VLMs through the whole process, from data to training and evaluation.
- Built WAON (AACL 2026 Findings), the largest native Japanese image-text dataset (155M pairs), and WAON-Bench. Noetra uses WAON to train its in-house vision encoder and WAON-Bench to evaluate Japanese cultural understanding.
- Built Jagle (COLM 2026; JSAI Annual Conference Award), the largest Japanese multimodal post-training dataset (9.2M examples). A 2.2B model trained on it beats InternVL3.5-2B on ten Japanese tasks, and Noetra uses it to post-train its multimodal LLM.
- Developed LLM-jp-4-VL 9B, whose beta matches Qwen3-VL-8B-Instruct on Japanese tasks.
- Built Japanese VLM evaluation: HakushoBench (AACL 2026 Findings), JAMMEval, and simple-evals-mm, a reproducible and reliable evaluation framework on 50+ benchmarks.
- Trained the speech language model Llama-Mimi (Best Paper Award, NLP2026) and llm-jp-modernbert.
Publications
Peer-reviewed
- 1
- 2
- 3
- 4
WAON: A Large-Scale Japanese Image-Text Dataset for Cultural Adaptation in Contrastive Vision-Language Models
Issa Sugiura, Shuhei Kurita, Yusuke Oda, Daisuke Kawahara, Yasuo Okabe, Naoaki Okazaki
AACL-IJCNLP 2026 Findings, Nov 2026
→ Used by Noetra to train its vision encoder; WAON-Bench evaluates its Japanese cultural understanding.
- 5
Jagle: Building a Large-Scale Japanese Multimodal Post-Training Dataset for Vision-Language Models
Issa Sugiura, Keito Sasagawa, Keisuke Nakao, Koki Maeda, Ziqi Yin, Zhishen Yang, Shuhei Kurita, Yusuke Oda, Ryoko Tokuhisa, Daisuke Kawahara, Naoaki Okazaki
COLM 2026 (acceptance rate 29%), Oct 2026
→ Used by Noetra to post-train its multimodal LLM.
- 6
- 7
- 8
- 9
- 10
- 11
Preprints
- 1
JFinBench: A Continuously Refreshable Japanese Financial Benchmark from the Latest Certification Exams
Issa Sugiura, Taro Makino, Takashi Ishida
arXiv, Nov 2026
- 2
- 3
- 4
- 5
- 6
LLM-jp: A Cross-organizational Project for the Research and Development of Fully Open Japanese LLMs
LLM-jp (incl. Issa Sugiura; 80 authors, listed in alphabetical order)
arXiv, Jul 2024
Domestic conferences (non-refereed)
- 1
- 2
Jagle: 視覚言語モデルのための大規模日本語マルチモーダル事後学習データセットの構築
全国大会優秀賞杉浦一瑳, 笹川慶人, 中尾圭佑, 前田航希, Yin Ziqi, Yang Zhishen, 栗田修平, 小田悠介, 徳久良子, 河原大輔, 岡崎直観
JSAI2026, 2026年6月
- 3
- 4
- 5
WAON: 視覚言語モデルのための大規模かつ高品質な日本語画像・テキスト対データセット
委員特別賞杉浦一瑳, 栗田修平, 小田悠介, 河原大輔, 岡部寿男, 岡崎直観
NLP2026, 2026年3月
- 6
Common Crawlを用いた大規模音声音響データセットの構築
浅井航平, 杉浦一瑳, 中田亘, 栗田修平, 高道慎之介, 小川哲司, 東中竜一郎
日本音響学会 2025年秋季研究発表会, 2025年9月
- 7
- 8
- 9
- 10
LLM-jp-3 VILA: 日本語マルチモーダルデータセット及び強力な日本語マルチモーダルモデルの構築
委員特別賞笹川慶人, 前田航希, 杉浦一瑳, 栗田修平, 岡崎直観, 河原大輔
NLP2025, 2025年3月
- 11
Awards
Honors
- Jun 2026
JSAI Annual Conference Award (66/1,376, top 4.8%)
Jagle: Building a Large-Scale Japanese Multimodal Post-Training Dataset for VLMs
JSAI2026 · Award list
- Mar 2026
Best Paper Award (3/789, top 0.4%)
Llama-Mimi: Speech Language Models with Interleaved Semantic and Acoustic Tokens
NLP2026 · Award list
- Mar 2026
Special Committee Award (32/748, top 4.3%)
WAON: Large-Scale and High-Quality Japanese Image-Text Pair Dataset for VLMs
NLP2026 · Award list
- Mar 2025
Special Committee Award (32/726, top 4.4%)
LLM-jp-3 VILA: Building a Japanese Multimodal Dataset and a Strong Japanese Multimodal Model (co-author)
NLP2025 · Award list
- Feb 2024
LINE Yahoo Sponsor Award (1/420, top 0.24%)
Improving Generalization of Trained Models via Unlearning of Mislabeled Data
DEIM2024 · Award list
Fellowships and scholarships
- Jul 2026
Full Repayment Exemption for Outstanding Achievement
JASSO graduate scholarship
Japan Student Services Organization
- Apr 2026
Science Tokyo Support Program for Doctoral Students
Doctoral research fellowship (JST SPRING)
Institute of Science Tokyo
Models and systems
- 1
LLM-jp-4-VL 9B series (2026)
Developed a 9B Japanese vision-language model extending the Japanese LLM LLM-jp-4 8B Thinking. To build this model, also constructed the training datasets Jagle, the largest-scale Japanese multimodal dataset, and RefinedVision, a refined version of FineVision addressing its license and quality issues.
- 2
simple-evals-mm (2026)
Developed a reproducible and reliable evaluation framework for LLMs and VLMs on 50+ English and Japanese benchmarks, used by the LLM-jp Multimodal WG, with a public leaderboard.
Certifications
- Oct 19, 2025 TOEIC L&R: L 420 + R 425 = 845
- Aug 2022 Security Camp organized by IPA (Information-technology Promotion Agency), Web Security Course