東京科学大学 環境・社会理工学院 融合理工学系. クロス研究室
Institute of Science Tokyo, School of Environment and Society
Transdisciplinary Science and Engineering
Meet The Team
クロス研究室の教育工学グループ
クロス研究室の教育工学グループの学生は、メタ認知、個人向けオンライン適応学習システム(POALS)の開発、VR(仮想現実)支援型言語学習、AI支援型ライティング、小学生のプログラム教育、生涯学習、3D 360度ビデオによる学習者エンゲージメントの向上、アイトラッキングと画像など、様々なテーマに取り組んでいます。特に、教育工学と人工知能をどのように組み合わせれば、学生の学習を改善できるか、そして、これらの新しい手法の有効性を評価できるかに関心があります。
Luc Gougeon (LUC)
GEDES D3+ student
Luc is a third-year Canadian working adult doctoral student. His research focuses on educational policies and computational thinking. He is interested in understanding if Japanese in-service teachers are ready to teach programming in 2020. Luc has been living in Japan since 2008 and has been working as a university lecturer since 2015. Luc uses technology in his classes on a daily basis and hopes that his research will help him train the new generation of educators.
Promoting University Students and Elementary School Teachers to Become Lifelong Learners Through Play
In 2020, Japanese primary school educators will face the difficult challenge of introducing programming in their classes despite the fact that they never studied programming themselves. Our research aims are mapping the specific contours of the knowledge gap in-service teachers and extend this surveying to current universities students who are also lacking computer literacy skills. Most research in the field of computer literacy places a strong emphasis on children while neglecting the needs of in-service educators and older students. We will tackle this research by both surveying a range of students and teachers while conducting case studies consisting of an education intervention meant to give university students a quick grasp of computational thinking, computer literacy and basic programming concepts. The case study approach intends to offer students essentials skills in an active learning environment, skills which will be transferable to their future workplace or classroom if they intend to become educators. The results of this study are intended to offer stakeholders and policy-makers a clearer picture of the current educational landscape and enlighten their decisions. Below is an illustration of summarizing the issues which will be investigated related to education approaches and students’ knowledge needs.
Dong Wang (Don)
GEDES D3 student
Don is a doctoral student who joined Cross Lab in October 2024. As a former professional game developer, he is interested in educational games and aims to combine his professional experience with academia. Don is particularly focused on applying VR and AI to educational games to create innovative learning experiences.
Utilizing Virtual Reality in a Serious Game for Immersion based Learning Outcomes in Shokuiku (食育/Food Education)
Shokuiku (食育/Food Education) is uniquely promoted by Japan through established laws and systems. In March 2021, the "Fourth Basic Plan for the Promotion of Shokuiku" was launched to guide efforts from FY2021 to FY2025. To achieve the various goals outlined in the plan, innovative educational tools are essential, as traditional approaches often lack engagement.
Serious games, designed for purposes beyond entertainment, are emerging as effective tools in fields like education and healthcare by combining interactive gaming elements with instructional design. Immersive VR serious games utilize VR technology to create engaging environments, with studies showing their effectiveness in various learning contexts. Generative AI is increasingly being used in immersive VR serious games, enhancing their interactivity, realism, and adaptability by creating personalized content and adapting gameplay based on player performance. This personalized approach improves engagement and learning outcomes.
However, while immersive VR games have been widely studied for physical exercise, there is limited research on their use in food education. This study explores how an immersive VR serious game, enhanced with generative AI, can promote Shokuiku by improving nutritional knowledge and self-efficacy, addressing current research gaps.
Porama Rittibunlue (Song)
GEDES M2 student
Development and Evaluation of an LLM-Powered Virtual Teaching Assistant for Online Course Support
Since the emergence of powerful Large Language Models (LLMs) such as ChatGPT in 2022, education has undergone a profound transformation. Many students now rely on LLMs as learning aids to support their studies and complete assignments. At the same time, educators are increasingly expected to understand these technologies and integrate them into their teaching to enhance learning outcomes. However, this rapid transition also raises important challenges. Students may encounter inaccurate or misleading information and risk diminished critical thinking skills, while educators may face increased pressure and workload as they adapt their instructional approaches.
Recent studies have explored the development of Virtual Teaching Assistants (VTAs) that leverage LLMs combined with Retrieval-Augmented Generation (RAG) techniques to assist both teachers and students. These tools aim to reduce teachers’ workloads while providing students with accurate, accessible learning support. However, existing approaches often struggle to handle queries beyond their RAG knowledge base or remain confined to specific subject domains, limiting their generalizability. This study seeks to design a more flexible and adaptive VTA that empowers both educators and learners to effectively utilize LLMs for diverse educational purposes.
Hiroyuki Sato
GEDES M1 student
The Impact of Mind Wandering on learners' Gaze Patterns
Research Description
This study aims to detect mind wandering (MW) during learning using a non-invasive, low-cost method based on built-in laptop cameras, and to quantitatively clarify its impact on learning performance. In the experiment, participants watch an online lecture while their facial and gaze data are captured using a Python-based gaze estimation system (Python-Gaze-Face-Tracker). Every 30 seconds, participants self-report whether they were attentive or not, providing binary MW labels. The collected parameters include head rotation angles (Pitch, Yaw, Roll), eye center coordinates, and pupil vectors, all sampled at 30 Hz. From these time-series data, statistical and dynamic features are computed for each 30-second window and analyzed in relation to MW labels. The analysis examines whether significant differences exist between attentive and MW states in gaze dynamics, head motion, and pupil variation, identifying features that effectively indicate MW.
Expected outcomes include: (1) identification of gaze-based indicators effective for detecting MW; (2) development and open release of a labeled gaze dataset and analysis scripts, establishing a reproducible research foundation; and (3) empirical demonstration of the feasibility of low-cost MW detection using standard webcams. Ultimately, this research contributes to the integration of educational technology, cognitive science, and AI-based learning support, enabling adaptive interventions and attention monitoring in online learning environments.
Zhang Min
ESI M1 student
Evaluating Quality–Safety–Energy Trade-offs in Local SLM-RAG for Heart-Failure Discharge Education
Local small language models (SLMs) may provide resource-efficient access to heart-failure discharge information, but their answer quality, clinical safety, and energy cost remain uncertain. This research focuses on simulated patient questions about heart-failure discharge education and follow-up care. The study compares three system configurations: direct answering with a local SLM, retrieval-augmented answering using permitted patient-education sources, and retrieval followed by an additional safety-control stage. The systems will be evaluated for medical correctness, clinical safety, and whole-system energy use. The goal is to examine the trade-offs introduced by retrieval and safety control under the same local model and computing environment.
Mona Guo
GEDES M1 student
Evidence-Constrained LLM Support for Longitudinal Analysis of Teacher Observations of Students with ASD
Teachers supporting students with autism spectrum disorder (ASD) may record observations across many days and classroom situations. These notes can be incomplete or contradictory, making it difficult to identify patterns without overlooking important exceptions. Large language models (LLMs) can help summarize the records, but they may also present interpretations as facts or draw conclusions that the notes do not support. This research aims to develop a teacher-facing framework that organizes observations over time, links analytical claims to their supporting records, and makes counterevidence and uncertainty visible. Using synthetic longitudinal classroom cases, the study will compare direct LLM analysis, prompt-optimized analysis, and the proposed evidence-constrained approach. Teachers in Japan will also evaluate selected outputs for clarity, trustworthiness, and usefulness for reflection. The system is intended to support teachers’ professional judgment, rather than diagnose students or make educational decisions for them.
Mingduo Li
ESI M1 student
Evaluating Interactive Dynamic Visualization in LLM-Supported Self-Learning of Two-Dimensional PCA
Students learning linear algebra may be able to perform calculations without fully understanding their geometric meaning. Two-dimensional principal component analysis (PCA) illustrates this challenge: learners need to connect covariance, eigenvectors, basis change, and projection to understand how principal directions are determined and what information is lost through dimensionality reduction. Large language models (LLMs) can provide conversational explanations, but learners may still find it difficult to understand how these mathematical relationships change when the data or projection direction changes.
Guided by multiple-representation theory, this research will evaluate whether interactive dynamic visualization supports university students’ conceptual understanding of two-dimensional PCA during LLM-supported self-learning. The study will compare two learning conditions: LLM explanations accompanied by matched static figures, and equivalent explanations accompanied by interactive dynamic visualizations created with GeoGebra. Both conditions will use the same mathematical content, datasets, labels, formulas, and matched key states, with common mathematical checks applied to the LLM support. In the interactive condition, learners will be able to modify data points and explore changes in covariance, principal directions, coordinate representations, projection, and information loss. Learning outcomes will be assessed using pretests and posttests, while questionnaires will examine perceived usefulness, ease of use, and engagement. The study aims to provide evidence on the educational value of adding dynamic interaction to otherwise comparable LLM-supported learning materials.
Jiyu CHEN
ESI M1 student
Evaluating Verification-Guided LLM Feedback in Introductory Data Analysis
Running Python code is one thing; understanding what its output means is another. In introductory data analysis, students must learn to choose statistical methods, examine their assumptions, and turn numerical results into meaningful conclusions. Large language models (LLMs) can offer immediate feedback, but they may misidentify mistakes, give unsupported explanations, or reveal answers too early. Can grounding their feedback in independently verified evidence help students understand their mistakes and tackle new problems?
Guided by Cognitive Load Theory, this research will compare direct LLM feedback with verification-guided feedback in university students’ Python data analysis learning. In the verification-guided condition, a separate rule-based module will check students’ submissions and supply structured evidence to the LLM. This evidence will guide targeted feedback that helps students revise their work without providing complete solutions.
Each participant will experience both conditions through matched but different task sets, with the order randomized and counterbalanced. Tasks will cover descriptive-statistics interpretation, hypothesis-test selection and assumptions, and test-result interpretation. Both conditions will use the same language model, interface, feedback structure, and target response length.
The main measure of learning will be improvement from initial to revised responses. Additional tasks completed without feedback will assess whether students can apply what they have learned to similar new problems. Questionnaires will examine mental effort, perceived usefulness, and trustworthiness. The study asks whether independent verification can make LLM feedback more educationally effective—not only helping students correct an answer, but also helping them reason through the next problem.
.png)