Guiding (digital) learning paths: Intelligent tutoring systems

In education, the synergy between technology and pedagogy has led to new approaches that allow for a reimagining of traditional classroom practices. Artificial Intelligence (AI) is becoming increasingly present in this context, but its use in education varies greatly and is far from widespread. One concern often voiced in discussions around AI is that such systems might make decisions that are opaque and difficult to understand. To address this concern, it is essential to understand how such systems work. This article introduces Intelligent Tutoring Systems and explains how learning can be personalized using these systems. Through personalization, learners can follow individual learning paths while pursuing the same learning goals. Intelligent Tutoring Systems can enable personalized learning even in large and heterogeneous learning groups, supporting teachers in creating more individualized learning experiences.

Intelligent Tutoring Systems

Intelligent Tutoring Systems (ITS) have been used since the 1980s to support learning in various application domains. Particularly in the United States, Intelligent Tutoring Systems have been developed and implemented in schools for subjects such as physics (e.g., AutoTutor, Graesser et al., 2005) and mathematics (e.g., MathTutor, Carnegie Mellon University, 2009-2023). Numerous large-scale studies in the U.S. have demonstrated that the use of Intelligent Tutoring Systems can improve learner performance (e.g., Kulik & Fletcher, 2016; VanLehn, 2011).

A core feature of these systems is their ability to provide each learner with individualized, tailored feedback or learning content based on an analysis of the learner’s progress (Issing & Klimsa, 2002). This adaptability is crucial because learners have different support needs, which can also change on an individual level throughout the learning process. Tutoring systems can address various aspects, including knowledge, strategies, self-regulation, or motivation (Aleven et al., 2017). The support is provided “just in time,” i.e., at the exact moment it is needed during the learning process. The idea is that learners learn best when they receive immediate feedback from a tutor (learning companion). This can be in the form of hints, such as: “You have constants on both sides of the equation. How can you arrange them so that all constants are on the left side and none on the right?” Feedback can also follow an incorrect input, such as: “A equals sign is missing. Make sure to enter the entire equation.” (See Figure 1.)

The goal is not to replace teachers but to enable one-on-one feedback where personalization and differentiation would otherwise be difficult—such as in large groups or during practice at home. In these contexts, teachers can use Intelligent Tutoring Systems to implement individual learning while keeping track of progress, allowing them to intervene pedagogically and didactically when necessary.

Learner models as the basis for tutoring decisions

When using Intelligent Tutoring Systems, learners work on tasks that involve sequential steps. As they progress through these steps, learners receive automatic feedback on the correctness of their inputs, such as calculations or answers. Additionally, the system can provide hints on request or automatically, explaining how to solve a problem correctly or how to correct mistakes.

The key to tailored feedback lies in modeling learner behavior, referred to as the learner model. Based on the learner model, the system can estimate which knowledge components have already been mastered and which still need to be learned. A typical Intelligent Tutoring System includes not only the learner model but also an expert model, a tutor model, and an interface (Nwana, 1990). The expert model represents the knowledge domain. It reflects the level of knowledge learners are expected to achieve by the end of their learning process and helps identify errors by comparing learner inputs with predefined rules. The tutor model provides the pedagogical basis for feedback, such as suggesting alternative learning strategies, and enables the system to select tasks or deliver targeted feedback based on the learner’s progress. The interface represents the learner’s view, displaying tasks in a didactically meaningful format, such as multiple-choice questions, simulations, or chat dialogues. 

Adaptivity of tutoring systems using the example of “Lynnette”

Using learner models and Machine Learning techniques, intelligent tutoring systems offer a dynamic framework of context-sensitive hints, assistance, and feedback that guide learners step by step through tasks. These systems analyze individual knowledge and its application to tasks to promote the development of specific knowledge components. An example of a freely available Intelligent Tutoring System for solving equations is Lynnette. Developed by researchers at Carnegie Mellon University in the U.S., Lynnette is accessible via MathTutor. Designed for mathematics instruction, Lynnette is a tutor for solving equations, offering step-by-step guidance, feedback on the correctness of solutions, error feedback, and adaptive problem selection.

Model-Tracing and Knowledge-Tracing

Two key technical methods enable Intelligent Tutoring Systems like Lynnette to adapt their responses to learner behavior: Model-Tracing and Knowledge-Tracing.

1. Model-Tracing tracks the learner’s individual approach to a problem and monitors their solution steps. If learners follow steps that suggest a common misconception, the tutor can respond with corrective feedback.
2. Knowledge-Tracing assesses the learner’s current level of understanding and evaluates their performance. Based on this assessment, the system adaptively selects new tasks. This involves monitoring progress—specifically, the mastery of certain knowledge components—through interactions with learning tasks. The system then selects tasks that align with the learner’s current level and pace, adjusting the difficulty or revisiting or skipping tasks as needed. This ensures learners receive the appropriate level of practice for all knowledge components covered by the tutor.

Outlook: Intelligent Tutoring Systems and the challenges of personalization

Intelligent Tutoring Systems enable learners to pursue individualized learning paths while aiming for common learning goals. By analyzing and tracking learning paths, educators gain insights into the learning process. This is particularly valuable for large and/or heterogeneous groups and for practicing at home, where personalized learning and differentiated instruction can be supported through the targeted use of intelligent tutors. For these systems to be effective, they must be evaluated to ensure alignment with curricula. This means these systems must be continuously reviewed and adjusted to ensure they adequately cover the defined sub-competencies in their respective domains. Although the development of such systems remains time- and cost-intensive, significant progress has been made in recent years. For example, researchers at Carnegie Mellon University, including the team behind Lynnette, have developed CTAT, an authoring tool that allows researchers to create Intelligent Tutoring Systems without requiring extensive programming skills. Hybrid forms of adaptivity are also emerging, combining various psychological aspects, such as emotions and motivation, with knowledge, learning paths, and strategies. However, evidence-based research remains essential to ensure that automation and AI are applied in education where they are didactically meaningful.

References