Research

TL;DR

Research Theme: Development of intelligent tutoring systems to optimize human thinking and learning

Central Challenge: Solving the trade-off between generalizability (applicable across various domains) and adaptability (optimized for specific domain characteristics)

Solution Approach: Creating “intermediate representations” that bridge human thinking and computer processing, enabling transfer of both support technologies and thinking skills across domains

Four Research Projects:

• CHUNK: Intermediate representation for knowledge transfer (three-layer model of function-behavior-structure)
• CLOVER: Intermediate representation for learning process optimization (mechanisms to learn effectively from errors)
• OCEAN: Intermediate representation for learning environment adaptation (integrating individual cognition, motivation, and goals)
• CCS: Intermediate representation for educational system design (common vocabulary enabling design knowledge sharing)

Expected Outcomes: Realization of truly generic learning support where support technologies developed in one system and thinking skills acquired in one domain can be leveraged in other domains

Research Overview

This page organizes my research projects and introduces the themes I am pursuing under the grand vision of developing intelligent tutoring systems that emphasize both generalizability and adaptability for optimizing human thinking and learning.

This vision begins with the ultimate question:

“What is needed for people to learn, think, and grow better?”

The most important challenge in addressing this question is the trade-off between generalizability and adaptability. For example, a learning system specialized for mathematics is effective for math but cannot be used for programming or language learning.

However, the problem extends beyond this. The inability to transfer technologies across domains also makes it difficult to support the acquisition and transfer of generic thinking skills, such as applying logical thinking learned in mathematics to programming or using structured thinking developed in programming for essay writing.

On the other hand, systems that can handle all domains struggle to provide fine-grained support that leverages the characteristics of individual domains or to foster deep thinking skills.

The core of my research lies in creating “intermediate representations (mediums)” that bridge these two aspects. In other words, by identifying common patterns in how humans learn and think and expressing them in forms that computers can process, I aim to realize systems that can be applied to various domains while providing support tailored to the characteristics of each domain. This will enable seamless learning support where learners can apply thinking skills and problem-solving approaches acquired in one domain to other domains.

This page presents the foundation of my research, central themes, and the projects that support them.

Research Foundation: Approach and Methodology

Why “Intermediate Representations” Are Important

Most conventional tutoring systems focus on statistically estimating learners’ understanding states and do not attempt to deeply understand how learners think and why they arrive at certain answers—the “content of thought”.

However, to achieve truly effective learning support, we need to:

1. Understand the essence of learners’ thinking processes: Grasp why they struggle and where their understanding becomes ambiguous
2. Design common mechanisms that work across different domains: Apply support methods effective in mathematics to programming and physics
3. Express in forms that computers can process: Convert complex human thinking into forms that systems can understand and utilize

“Intermediate representations” achieve these goals, serving as “translators” between human thinking and computer processing.

Design of Computable and Portable Frameworks

To optimize learning and thinking across domains, we need frameworks that are not overly abstract but concrete and practical. Specifically:

• Computable: Processable by computers and applicable to real-time learning support
• Portable: Mechanisms developed in one domain can be applied and adapted to other domains
• Concrete: Reflecting learners’ actual thinking processes and knowledge structures

For example, statistical information like “the learner’s comprehension level is 0.7” makes it unclear how to support that learner effectively. Instead of such approaches, it is important to enable systems to handle concrete and meaningful information such as “this learner understands the concept of variables but is confused about loop processing” or “they can do basic calculations but cannot formulate equations for word problems”.

Three-Layer Research Approach

My research is realized through the following three-layer structure:

Human Thinking/Learning → Intermediate Representation (Translation/Conversion) → Computer-based Support

1. Understanding Human Thinking/Learning: Detailed analysis of how learners think and learn through cognitive science, educational psychology, and learning sciences
2. Design of Intermediate Representations: Development of mechanisms to convert human thinking into forms processable by computers while preserving meaning
3. System-based Support: Automatic provision of personalized learning support using intermediate representations

Key Research Elements

Elucidation of Thinking Processes: Detailed analysis of how learners understand problems, formulate solutions, and execute them.

Design of Knowledge Structures: Organization and structuring of learning content in forms that are easily understandable for learners and processable for systems.

Optimization of Learning Activities: Dynamic design and provision of the most effective learning activities according to learners’ cognitive characteristics and understanding situations.

Intelligent Tutoring Systems: Development of systems that integrate the above insights to provide real-time, personally optimized learning support.

By integrating these elements, I aim to create an environment where learners can focus on essential learning.

Related Fields and Their Integration

My research spans interdisciplinary fields such as “cognitive science” and “knowledge engineering.”

I will introduce how each field relates to the design of intermediate representations through several keywords:

• Science: Theories for understanding the essence of human cognition and learning (e.g., cognitive science, cognitive psychology, educational psychology, learning sciences, psychophysics)
• Engineering: Technologies for implementing and utilizing intermediate representations (e.g., knowledge engineering, learning engineering, applied ontology/ontology engineering, artificial intelligence, computer science, educational technology, human-computer interaction)
• Philosophy: Design principles and conceptual foundations for intermediate representations (e.g., constructivism, representationalism, operationalism, model-driven, analysis-by-synthesis)

Research Projects

My research builds upon “Generalization of Intelligent Tutoring Systems Based on Information Structure-Oriented Approach” established in my doctoral dissertation. In my doctoral thesis, I proposed new methods for applying the “information structure-oriented” approach—which segments learning content into components and describes them in computer-readable states—to immature learning domains, as well as methods for reusing systems based on this approach and extending them to support higher-order abilities.

Current research further develops this foundation by capturing essential patterns of human thinking and learning as “intermediate representations” to explore new approaches that solve the trade-off between generalizability and adaptability. The following four projects each concretize the concept of “intermediate representations” from different angles and contribute to my research vision.

CHUNK: Intermediate Representation for Knowledge Transfer

(Componentization of Human Understanding and Knowledge)

Research Background and Essential Problem

There is a problem in learning where even after acquiring skills, learners cannot apply them in other situations. For example, understanding “for loops” in programming but being unable to use them in actual problem-solving, or memorizing proof methods in mathematics but being unable to use them in application problems.

The essence of this problem lies in learners remaining at the level of superficial memorization of procedures without understanding why those procedures work (behavior). In my doctoral research, approximately 80% of learners could not apply acquired subgoals (reusable procedures) to new problems, which became the starting point of this research.

Solution Approach Through Intermediate Representation

The CHUNK project develops intermediate representations that express knowledge in three layers: “function-behavior-structure”. This builds upon the information structure-oriented approach established in my doctoral thesis and has the following characteristics:

• Function: What the knowledge is used “for” (purpose)
• Behavior: “How” the knowledge operates (execution process)
• Structure: “What kind of” procedures the knowledge consists of (implementation)

For example, the subgoal “array processing” in programming includes:

• Function: Apply the same processing to multiple data items
• Behavior: Access each element while sequentially changing the index
• Structure: for loop + array access + processing content

This information is structurally contained.

Specific Research Themes

BROCs (Building method that Realizes Organizing Components)

A stepwise expansion method for components in knowledge organization in programming. It enables learners to structurally accumulate components step by step through two repeating steps: (1) combining basic elements (assignment statements, if statements, etc.) to create small functions (swap, etc.), and (2) combining already created components to create larger functions (sort, etc.).

BEAR (Educational Behavior Analyzer of Source Code for Understanding Behavior Model)

An educational behavior analyzer that supports learners’ understanding of source code behavior. It receives source code and dataset generation rules, generates random datasets using rule-based methods, analyzes and visualizes variable state changes corresponding to each line execution by applying generated datasets to source code, supporting learners in understanding behavior models.

Compogram

Integrates the above mechanisms to provide a learning environment where learners can gradually acquire subgoals while understanding “behavior.” It combines BROCs’ stepwise expansion method with BEAR’s behavior analysis function to support knowledge acquisition based on deep understanding rather than superficial memorization. This research proposes the acquisition of “Subgoal Flexibility” — the ability to apply component knowledge learned in one problem-solving context to new situations.

Relationship to Vision

The CHUNK project directly contributes to the core of the research vision: solving the trade-off between generalizability and adaptability.

Conventional programming education systems specialize only in programming, making it difficult to apply thinking skills acquired there to other domains. However, the intermediate representation of “function-behavior-structure” developed by CHUNK shows potential for transferring not only programming but also “thinking that decomposes problems into components and assembles them step by step” to any situation requiring structural thinking, such as constructing mathematical proofs, logical composition in writing, and designing research plans.

This research enables learners to apply learning from one domain to other domains, realizing true transfer of thinking skills. This is an important theme that forms the foundation of my research vision of “optimizing human thinking and learning.”

CLOVER: Intermediate Representation for Learning Process Optimization

(Computational Learning Optimization with Variform External Representations)

Research Background and Essential Problem

“Errors” and “failures” in learning are often considered things to avoid, but errors actually contain valuable learning information, revealing learners’ thinking processes and ambiguous understanding. However, conventional systems remain at the level of “correct/incorrect” judgments and cannot provide effective support for learning from errors.

Solution Approach Through Intermediate Representation

The CLOVER project develops a “transformation model” that adaptively converts learners’ responses into effective external representations as its intermediate representation. This is based on “Error-visualization” technology that has been continuously researched for over 20 years, combined with the concept of Multiple External Representation (MER) in an integrated research framework.

The core of error visualization lies not in providing direct feedback or correction to learners’ responses, but in selecting and converting to optimal visualization representations from diverse external representation options (language, diagrams, mathematical expressions, tables, graphs, etc.), enabling learners to spontaneously realize “this is different from what I expected” and “this is different from the correct answer,” and motivating deep knowledge revision through cognitive conflict.

Therefore, CLOVER’s intermediate representation, the “transformation model,” includes the following three functions:

• Adaptive Selection: Selecting optimal external representations according to the nature of learners’ responses
• Effective Transformation: Generating visualization representations that easily induce cognitive conflict
• Spontaneous Correction: Promoting learners’ internal error correction rather than direct feedback

This transformation process is expected to provide optimal visualization support tailored to learners’ understanding situations, even for the same “wrong answer.”

Specific Research Themes

EBS (Error-based Simulation)

A simulation-based learning environment using error visualization that has been continuously researched for over 20 years. It converts learners’ responses into simulation representations showing “phenomena that would occur if the learner’s answer were correct,” promoting learning through trial and error. Research particularly focused on mechanics adaptively presents auxiliary problems in stuck situations, supporting learners in recognizing errors themselves.

TAME (Teachable Agent Modeling for Error-visualization)

An approach that utilizes Teachable Agents as “simulators for hypothesis testing by learners” and deepens learning using error visualization technology. By inputting “rules to be learned in learning tasks,” it aims to realize control that enables learners to interpret and explore the behavior of Teachable Agents.

ELMER (Explainable Model for Learning from Errors with Multiple External Representations)

A framework for designing observable feedback to support exploratory learning in acquiring hard-to-observe concepts. It converts abstract concepts into multiple observable external representations (Multiple External Representations) to promote error recognition and correction, enabling broad application of exploratory learning.

Relationship to Vision

The CLOVER project aims to realize the optimization of learning processes aspect of the research vision.

In conventional learning support, errors were treated as “failures” to be avoided, and learners tended to fear errors. However, through CLOVER’s transformation model, errors become utilized as “treasures of learning.” Unexpected results in physics experiments, calculation mistakes in math problems, bugs in programming, logical breakdowns in writing—all errors are converted into opportunities to deepen learners’ understanding.

This “thinking pattern of learning from errors” is a fundamental skill common to all creative activities, such as hypothesis testing in scientific research, prototype improvement in design thinking, and trial and error in artistic creation. Through CLOVER, learners are expected to acquire high-order learning abilities of challenging without fearing errors, learning effectively from failures, and continuously improving.

OCEAN: Intermediate Representation for Learning Environment Adaptation

(Optimizing Cognition by Engagement of Agent’s Navigation and Negotiation)

Research Background and Essential Problem

Modern learning environments are information-rich, leaving learners confused about judgments such as “what should I learn,” “in what order should I proceed,” and “can I move forward with my current understanding level.” Additionally, motivation for learning fluctuates, swinging between “I want to do this” and “this is troublesome.”

Conventional systems remained at fixed learning sequences or superficial progress management, unable to provide flexible adaptation according to individual cognitive characteristics and motivational states.

Solution Approach Through Intermediate Representation

The OCEAN project develops intermediate representations that integrate learners’ situations in multiple layers.

In modern learning environments, while vast amounts of information and resources are available, learners face complex judgments such as “where should I start,” “in what order should I proceed,” and “can I move forward with my current understanding level.” Furthermore, motivation for learning constantly fluctuates, swinging between “I want to do this” and “this is troublesome” for the same task depending on time and circumstances.

OCEAN’s intermediate representation uses a four-layer structure to express learners’ situations for integrated handling of these complex factors:

• Cognitive Layer: Which concepts are understood and where ambiguities exist (detailed grasp of knowledge states)
• Behavioral Layer: What learning behaviors are taken and which resources are used (analysis of learning patterns)
• Motivational Layer: Motivation for learning and its fluctuation factors (modeling of motivational states)
• Goal Layer: Relationship between short-term tasks and long-term goals (understanding of goal structure)

This aims to comprehensively understand each learner’s cognitive characteristics, learning style, motivational patterns, and goal setting, dynamically configuring personally optimized learning environments.

Specific Research Themes

WHALE (Wise Helper Agent for Learning Environment) + ARK (Action-Resource-Knowledge Model)

An educational agent that provides learning path recommendations in learning environments integrating multiple educational tools. The ARK model provides comprehensive recommendations for “what to learn,” “what materials to use,” and “what to do with those materials” through a multi-layer background model consisting of action sequence layer, resource item layer, and knowledge graph layer. It enhances recommendation acceptance through personalized educational agents.

CORAL (Cognitively-Recalibrated Adaptive Learning)

A new motivational framework focusing on the emotion “troublesome (Demotivated)” that cannot be explained by conventional motivation research. It calculates the degree of Willingness based on the balance between Motivation and Demotivation, analyzing and supporting learners’ internal motivation through interactive interactions.

Relationship to Vision

The OCEAN project contributes to the ultimate goal of realizing personally optimized learning environments in the research vision.

In modern society, information overload causes learners to experience confusion and bewilderment from having too many choices. Additionally, motivation for learning is complex and volatile, with emotional barriers such as “I don’t feel motivated today” or “this task is troublesome” hindering learning. OCEAN’s multi-layer integrated intermediate representation aims to comprehensively understand these complex factors and realize personally optimized learning support.

This “thinking pattern of designing and continuing optimal learning for oneself” extends beyond school education to demonstrate power in all aspects of lifelong learning, such as strategic preparation for certification exams, efficient acquisition of new technologies, improvement in hobbies and sports, and even life planning and career development. Through OCEAN, learners are expected to always select optimal learning strategies according to changing environments and their own situations, realizing continuous growth.

CCS: Intermediate Representation for Educational System Design

(Computational Cognitive Schemas)

Research Background and Essential Problem

In intelligent tutoring system research, even when different systems handle common problem-solving skills, linguistic and structural foundations for comparing and referencing them do not exist. Additionally, connections between task design and learning goals are rarely made explicit, leaving little basis for examining whether tasks truly contribute to acquiring targeted skills.

The essence of this problem lies in the absence of a “common vocabulary” that abstractly describes the design intentions of intelligent tutoring systems and enables comparison and reuse of problem-solving skills common across different systems. This causes fragmentation of design knowledge from individual systems, making integration of knowledge in learning sciences as a whole difficult.

Solution Approach Through Intermediate Representation

The CCS project develops intermediate representations as common vocabulary for abstractly describing the design intentions of intelligent tutoring systems. This is an approach that separates problem-solving skills from individual implementations and expresses thinking processes as sequences of operations and states.

While conventional educational system evaluation relied on quantitative evaluation of educational effects, CCS provides a framework for re-examining the validity of support from a design theory perspective. CCS intermediate representations have the following functions:

• Task Alignment Evaluation: Evaluating structural alignment between thinking processes promoted by learning tasks and target skill structures
• Inter-system Comparison: Comparing skill descriptions across systems with different domains or task formats
• Design Knowledge Sharing: Cross-domain reuse of support methods among systems with isomorphic CCS
• Meta-scientific Integration: Structurally connecting and reorganizing knowledge in educational system informatics

This framework enables sharing design principles beyond task formats, even when skills like “extracting and reconstructing structures” appear across programming education and writing composition support.

Specific Research Themes

CCS (Computational Cognitive Schemas) Common Vocabulary Framework

A common vocabulary that abstractly describes the design intentions of intelligent tutoring systems and enables comparison and reuse of problem-solving skills across different systems. By expressing thinking processes as sequences of operations and states, it supports cross-domain utilization of design knowledge beyond domains and task formats.

Design Knowledge Sharing Method Through Isomorphic CCS Identification

A method that organizes systems with different appearances or task contents but aiming to acquire common cognitive schemas as isomorphic CCS, realizing cross-domain reuse of support methods. It enables sharing design principles across fields, such as programming education and writing composition support.

Relationship to Vision

The CCS project contributes as a foundation for realizing solving the trade-off between generalizability and adaptability in the research vision at a meta-scientific level.

In conventional educational system research, knowledge from individual practices was fragmented, making accumulation and reorganization of design knowledge across fields difficult. Through the common vocabulary provided by CCS, intermediate representation technologies developed in CHUNK, CLOVER, and OCEAN projects can be described integratively, expected to realize construction of formal theoretical knowledge in educational system informatics.

This outcome will become the foundation for systematic development of intelligent tutoring systems based on design knowledge in the AI era, enabling scientific design of more effective learning environments utilizing design knowledge accumulated beyond individual systems.

Future Prospects: New Possibilities in the AI Era

Currently, with the emergence of Large Language Models (LLMs), new possibilities have emerged for fundamentally solving the previous trade-off between generalizability and adaptability.

Integration of LLMs and Intermediate Representations

The intermediate representations I have constructed through my research can expect new developments through integration with LLMs as follows:

• Automatic Knowledge Conversion: LLMs can automatically convert intermediate representations effective in one domain into forms applicable to other domains
• Dynamic Personal Optimization: Real-time adjustment of support based on intermediate representations according to changes in learners’ situations
• Natural Language Interpretability: Explaining the content of intermediate representations and rationale for support in natural language that learners can easily understand

Contribution to Society

In the AI era, humans are required to have higher-order abilities such as creativity, critical thinking, and problem-solving skills. The intermediate representation foundation constructed through my research will contribute to realizing a society where each individual can maximize their potential and continue learning throughout their lives.

Particularly, intermediate representation technologies developed along the four axes of thinking skill transfer, learning process optimization, individually adaptive learning, and educational system design are expected to be utilized in various learning contexts, bringing fundamental improvements to human learning abilities.

Acknowledgements

This research emerged from discussions with many colleagues, students, and collaborators. I am particularly deeply grateful to the following mentors who guided me: Takahito Tomoto, Tsukasa Hirashima, Tomoya Horiguchi, Hiroaki Ogata, Izumi Horikoshi, Rwitajit Majumdar, H. Ulrich Hoppe, Riichiro Mizoguchi, and Takako Akakura.