JA EN

Research Theme

Software

We study how apps and web services are built and look for better ways to develop them. The 3rd-year seminar and the development seminar also cover this area through team development and app building.

Test case generationCode review supportUser story writing support
Overview

What we work on

  • Learn the whole flow from deciding what to build to design, implementation, and testing.
  • Build games, student-support apps, AI apps, and more while sharing roles within a team.
  • Study what generative AI makes easier in development, and where care is needed.
Papers

Papers

Software Engineering Education

Assessing PBL work in the LLM era

A Year-over-Year Comparison of Artifact Size and Quality in Individual Free-Topic PBL Assuming LLM Use

This study analyzes how students' work in a free-topic PBL course changed once generative AI was available, focusing on a third-year undergraduate software engineering course.

LLMPBLSoftware engineering educationArtifact assessment
View paper (PDF)
Show details Hide details

Problem

In LLM-assisted development exercises, the size and quality of what students build depend not only on their skills but also on how they use LLMs and how the course is designed. We need to clarify what can actually be assessed from the deliverables.

Approach

We compared the functional size of deliverables from 48 students in 2024 and 46 students in 2025, using LLM-assisted measurement based on the COSMIC method. For 2025, we also examined the relationship between basic proficiency (measured by CFRP) and test pass rates.

Key results

Average functional size in 2025 was about 3.2 times that of 2024. CFRP showed only a weak correlation with functional size, but a moderate positive correlation with test pass rates. The study also states the limitations of LLM-assisted assessment.

Ongoing Research

Ongoing Research

Research topics the lab is working on as of the 2026 academic year.

An agile process for software development with generative AI: proposal and practice

We bring methods for reducing cognitive debt into an agile process and evaluate how development time and understanding of the software change.

Background

Generative AI code editors make it easy to put off understanding the code. This "comprehension debt" builds up until developers no longer understand how their software works, which we call "cognitive debt."

AgileGenerative AICognitive debt

An empirical analysis of quality and understanding in vibe coding

We measure prompt quality during development, the quality of the resulting app (via static analysis), and the author's understanding (via a comprehension test), and analyze how they relate.

Background

Vibe coding, where AI writes code from prompts and you move on as long as it runs, is fast, but raises concerns about code that is hard to fix and that its own author cannot explain.

Vibe codingGenerative AICode quality

Cognitive debt in software engineering with generative AI

Using four quadrants of subjective self-assessment and objective understanding, we clarify how cognitive debt arises and how severe it is, and explore software engineering education in which learners can use AI while keeping track of their own understanding.

Background

If you keep using AI-generated code without understanding it, you eventually become unable to fix or improve it. We call this accumulating lack of understanding "cognitive debt."

Cognitive debtSoftware engineering educationGenerative AI

A framework for evaluating and improving lab websites for promotion

We measure the usefulness of a lab website with indicators such as understanding of the research, intention to apply, and how well career paths and achievements come across, and feed the results into continuous improvement.

Background

Most lab websites have no mechanism for continuously measuring and improving whether they actually work as promotion.

WebsitesEvaluation metricsContinuous improvement

A generative AI system to support academic presentations in English

We are developing a web app that combines CEFR assessment, English script generation fitted to the talk length, and Q&A practice based on the slides and paper into a single workflow.

Background

For students who are not confident in English, the unpredictable Q&A after a talk at an international conference is a major burden. Existing AI tools focus on script generation and pronunciation scoring, and offer little support for practicing Q&A grounded in the research itself.

Generative AIEnglish educationWeb app

Automatic proofreading of novel dialogue considering persona and context

We study a dialogue proofreading feature that takes into account each character's persona (personality and habits) and the context. Existing dialogue from the target work is used as training data.

Background

Immersion in the world of a story matters in novels and game scenarios, but keeping characters consistent is hard in long works and in the early stages of planning.

LLMNatural language processingCreative writing support

Supporting people with autism using wearable devices

We combine biosignals from wearable devices (heart rate and sweating) with a self-exciting model to predict panics in advance and notify family members and caregivers, preventing dangerous behavior before it happens.

Background

Some people with autism may become violent toward caregivers or harm themselves during a panic. Predicting a panic in advance through human observation alone is difficult.

WearablesBiosignalsPrediction models

Contact

+81-47-469-5709

matsuno.yutaka(at)nihon-u.ac.jp

* To prevent spam, "@" is written as "(at)". Please replace (at) with @ when sending.

Matsuno Lab, Department of Computer Engineering, College of Science and Technology, Nihon University

Room 243, 4F, Building 2, 7-24-1 Narashinodai, Funabashi, Chiba 274-8501, Japan

Directions