Yato Shoma

About

I am a first-year master’s student at Tokyo Polytechnic University and a Research Assistant at the National Institute of Informatics (NII), LLM Center. My current research focuses on memorization and forgetting in LLM pretraining. My broader interests include generative models, world models, reinforcement learning, applied AI systems, and creative work.

Publications & Presentations

2026

• Hierarchical World Models Using Temporal Abstraction Based on Boundary Detection Official Japanese title: 「境界検出に基づく時間抽象化を用いた階層型世界モデル」 Daiki Murakawa, Shoma Yato, Shunsuke Sakai, Tatsuhito Hasegawa, Masahiro Suzuki, Yutaka Matsuo
  The 40th Annual Conference of the Japanese Society for Artificial Intelligence (JSAI 2026)
  G Messe Gunma + Online · June 8–12, 2026 · Session ID: 1E3-OS-39a-03 · DOI: 10.11517/pjsai.JSAI2026.0_1E3OS39a03
• Continual Pretraining and Analysis of a Japanese LLM Using a Government Publications Corpus Official Japanese title: 「官庁出版物コーパスを用いた日本語 LLM の継続事前学習とその分析」 屋藤翔麻 (東京工芸大), 清丸寛一 (NII), 小田悠介 (NII), 河原大輔 (早稲田大学)
  The 267th Natural Language Processing Research Meeting (IPSJ SIG-NL)
  March 7, 2026 Japanese venue name: 第267回自然言語処理研究発表会

2025

• Structured Memory Enhancement for Generative Agents Using Knowledge Graphs Final research presentation · Matsuo Lab LLM Community “LLMATCH” Cohort 2 · Sep 26, 2025 Slides · Recording
• Improving Exploration Efficiency in Sparse-Reward Environments via Refined Intrinsic Motivation and Contrastive Learning
  Hiroki Sagara, Shoma Yato, Shuichi Kusaba, Yousun Kang
  The 40th International Technical Conference on Circuits/Systems, Computers and Communications (ITC-CSCC 2025)
  July 8, 2025 Conference paper · WEIE Workshop · Related work from the same research project as the JSAI 2025 paper below
• Enhancing Exploration Efficiency in Sparse Reward Environments through Refined Intrinsic Motivation and Contrastive Learning Official Japanese title: 「内発的動機付けと対照学習の改良によるスパースな報酬環境における探索効率向上」 Hiroki Sagara, Shoma Yato, Shuichi Kusaba
  The 39th Annual Conference of the Japanese Society for Artificial Intelligence (JSAI 2025)
  Osaka International Convention Center + Online · May 27–30, 2025 · Session ID: 2Win5-14 · DOI: 10.11517/pjsai.JSAI2025.0_2Win514 Conference proceedings paper · Final project for the World Models / Deep Learning Applications Course, Matsuo–Iwasawa Laboratory, UTokyo

Selected Projects & Works

AI Development

2026

• JSAI 2026 40th Anniversary Special Session 6: AIWolf Contest 2026 Domestic Tournament Japanese official name: 40周年記念企画6：人狼知能コンテスト2026国内大会 2026 Spring Domestic Tournament, Natural Language Processing Division Japanese official division name: 2026春季 国内大会 自然言語処理部門 yatolab: “Improving an AIWolf Agent through Character Assignment and Separated Self/Other Models” Japanese project title: yatolab: キャラクター性付与と自己・他者モデル分離による人狼知能エージェントの改良 Co-developed yatolab’s Japanese-language AIWolf agent using character conditioning and separate self/other models for social-deduction dialogue. Presented the three-person project at JSAI 2026 as one of 19 final-round teams.

2025

• Voice Imitation AI App: VALL-E-X_JP-Voice-Cloner Built and deployed a Japanese zero-shot voice-cloning app from 3–5-second reference clips. Focused on the Gradio UI, Hugging Face Spaces / ZeroGPU deployment, and dependency fixes; shipped the public demo in under a week.
• A Lightweight ONNX-Based Toolkit: Waifu-GAN-Flow Built an ONNX / Gradio toolkit for full-body anime generation and latent-space animation, with seeded and batch generation, interpolation, CUDA inference, and MP4 export.
• An AI That Talks! Othello Commentary AITuber System — Tokyo Polytechnic University Festival 2025 Exhibition Project Team project integrating Egaroucid, an LLM API, VOICEVOX, Live2D / AITuberKit, and WebSocket for real-time spoken commentary. Containerized with Docker for remote GPU operation; the festival exhibit served 79 visitors. 工芸祭2025

2024

• The University of Tokyo May Festival 2024 — Virtual Cafe AI App Development 五月祭

Hackathon Projects

2025

• Tokyo AI Festival 2025 Hackathon Project: Disaster Support Web App Using a Local LLM Contributed to a multimodal disaster-support MVP that mapped text, voice, and image reports using Whisper, LLM-based extraction and triage, CLIP with LoRA, XGBoost, Supabase, geocoding, and real-time map visualization.

2024

• What-If Story Generator — A “What if…?” Prompt-to-Storytelling Assistant Powered by RAG and Gemini 1.5 Pro Created a capstone narrative assistant that turns a “What if…?” prompt into a structured Markdown story outline. Built a two-stage RAG pipeline with Gemini 1.5 Pro, embeddings, ChromaDB, and Google Grounding, plus retries, caching, and output checks.

Creative & Civic Works

• Cooperation for the 70th Anniversary Project of Atsugi City Administration (Colorization of Black-and-White Photos) 「厚木市政70周年記念事業協力」
• Atsugi City Administrative Content (2022) 厚木市行政コンテンツ2022
  • “Reunion in Atsugi — Atsugi OEC Food” Filming & Editing · 「厚木で再会～あつぎＯＥＣフード～」
  • “Architectural History of Atsugi — The History of Atsugi” Composition, Editing & Color Grading · 「厚木の建築史～厚木の歴史～」

Research & Work Experience

• National Institute of Informatics (NII), LLM Center — Research Assistant May 2026–present · リサーチアシスタント · Research on memorization and forgetting in LLM pretraining
• National Institute of Informatics (NII), LLM Center — Technical Assistant Dec 2025–Mar 2026 · 技術補佐員
• Matsuo Lab LLM Community “LLMATCH” (Cohort 2) — Community Researcher Apr 1–Sep 30, 2025 · Final research project on knowledge-graph-based structured memory for generative agents

Education

• Tokyo Polytechnic University, Graduate School of Engineering, Master’s Program in Engineering Apr 2026–present · 東京工芸大学大学院 工学研究科 工学専攻
• Tokyo Polytechnic University, Faculty of Engineering, Department of Engineering Apr 2022–Mar 2026

Awards & Honors

• AI and Semiconductors 2025 — High Achiever, Session 9 Exercise Matsuo–Iwasawa Laboratory, The University of Tokyo · 第9回演習 優秀生
• World Models 2024 — Outstanding Final Project Presentation Team recognition · Matsuo–Iwasawa Laboratory, The University of Tokyo
• Excellence Award for a Submitted Report in “Social Studies by Students” 「学生による社会スタディ」優秀証

Skills & Interests

• Research: LLM pretraining analysis, generative models, world models, and deep reinforcement learning
• AI & Web Development: CNNs, deep learning model implementation, ONNX, React, and Vue.js
• Infrastructure & Collaboration: Docker, AWS, Anaconda, and GitHub-based team development
• Creative Work: Illustration, music production (Hatsune Miku), 3DCG, and DJing

Selected Learning & Programs

Current

• Physical AI Fundamentals 2026 Matsuo–Iwasawa Laboratory, The University of Tokyo · In progress
• Large Language Models Course 1 (2026) Matsuo–Iwasawa Laboratory, The University of Tokyo · In progress

Selected Completed Programs

• World Models 2024 Matsuo–Iwasawa Laboratory, The University of Tokyo · Completed
• AI and Semiconductors 2025 Matsuo–Iwasawa Laboratory, The University of Tokyo · Completed
• Deep Generative Models 2025 Matsuo–Iwasawa Laboratory, The University of Tokyo · Completed
• Large Language Models Course 2025 — Advanced Matsuo–Iwasawa Laboratory, The University of Tokyo · Completed

Additional coursework includes deep learning, reinforcement learning, world models, LLM fundamentals, GCI, AI management, computer vision, TinyML / efficient deep learning, and AcademiX training programs.

Contact & Links

Email: m2663008@st.t-kougei.ac.jp