Hello everyone. I am Tanaka, a first-year doctoral student. In this article, I will introduce a paper I read at our recent English seminar.

•Paper Title: GPTalk: LLM-based virtual companions for metacognitive growth in self-regulated e-learning
•Authors: In-Taek Jung, ChungHa Lee, In-Chang Baek, Dongik Oh, Youjin Choi, Kyung Joong Kim, Duk-Jo Kong, Jin-Hyuk Hong
•Journal: International Journal of Human-Computer Studies, Volume 210 (2026)

This paper presents research utilizing a learning support system called “GPTalk” by Jung et al. (2026). The study proposes and evaluates a mechanism that supports learners’ metacognition in e-learning using LLM-based virtual “teachers” and “peers.”

1. Research Background
In e-learning centered on recorded lectures, learners often spend long periods studying alone, requiring them to check their own understanding and adjust their learning strategies independently. However, in practice, a common issue is that learners simply watch the videos and finish with a false sense of understanding. This study demonstrates the potential to address these issues by eliciting learners’ metacognitive activities through dialogue with LLM-powered virtual agents.

2. Learning Support System: GPTalk
GPTalk is designed as a system where virtual teachers and virtual peers interact with the learner at each stage of the learning process. RAG (Retrieval-Augmented Generation) is also incorporated to provide accurate responses based on lecture materials.

Preliminary Study for System Design
For the system design, this study first conducted semi-structured interviews with 10 high school students and 8 teachers to organize the needs of the educational field. From this, six design requirements were derived as follows:
1. Support for goal setting and planning
2. Promotion of self-monitoring
3. Support for adjusting learning strategies
4. Post-learning summarization and self-reflection
5. Motivation maintenance
6. Stimulation of metacognition through interaction with peers

Support for Each Learning Phase
•Before the lecture: The virtual teacher points out key aspects of the lecture and has the learner verbalize their learning goals.
•During the lecture: Comprehension check questions and quizzes are presented in the middle of the video. Furthermore, the virtual peer makes statements like “I don’t understand this,” creating an opportunity for the learner to explain, which encourages the reconstruction of understanding.
•After the lecture: The learner reflects on the content through dialogue with the virtual peer, and the virtual teacher intervenes as necessary to encourage the correction of misconceptions.
In this way, the system is designed not just for simple Q&A, but follows the flow of Self-Regulated Learning: goal setting → monitoring → reflection.

3. Research Methodology
Participants were 32 high school students in South Korea. In an economics lecture, a group using GPTalk was compared with a group using only simple Q&A chat with an LLM. Evaluation of learning outcomes, accuracy of self-assessment, and learning experience was conducted using the following methods:
•Learning outcomes: Measured understanding of learning content through pre- and post-tests.
•Metacognitive monitoring accuracy: KMA (Knowledge Monitoring Accuracy). An index that measures how accurately one can estimate whether their own answers are correct (= accuracy of metacognitive monitoring).
•Learning experience: UEQ (User Experience Questionnaire). A questionnaire evaluating user experience such as ease of use, clarity, and satisfaction with the system.

4. Results
1) Learning outcomes (test scores)
Although a trend toward higher scores was observed in the condition using GPTalk, it was reported that this did not reach statistical significance. However, considering that this was a short, single learning session, it is interpreted as meaningful that a trend toward promoting rather than hindering learning was observed.

2) Metacognitive monitoring accuracy (KMA)
What is particularly important in this study is the result regarding the suppression of the “illusion of understanding.”
•In the baseline group (the group not using GPTalk), there was a tendency for instances of “being confident” despite incorrect answers to increase. On the other hand, in the GPTalk group, this increase was suppressed.
•Also, in the GPTalk group, the accuracy of “correctly recognizing” when they did not understand improved.
In other words, it is suggested that using GPTalk made it easier for learners to grasp more accurately “what they understand and what they do not understand.” This is considered a noteworthy result as support for enhancing not only the amount of knowledge itself but also the accuracy of self-assessment (calibration).

3) Learning experience: User experience and qualitative evaluation (UEQ)
In the UEQ, GPTalk received significantly higher ratings in multiple items. The following points were particularly characteristic:
•Social presence
Due to the presence of the virtual peer, the “feeling of studying alone” was weakened, which served as support for continuing to learn.
•Learning by teaching
It was reported that the process of the learner correcting the virtual peer’s errors most stimulated metacognition.
Based on these results, it is stated that for support using LLMs, it is important to design not just to “provide answers,” but to have learners explain and think.

5. Discussion
LLM-based companions and metacognition
In this study, GPTalk acts as a mechanism to promote metacognition by having learners verbalize goal setting and summaries. Furthermore, because the LLM can provide context-aware responses according to the learner’s input, it is said that more natural and flexible scaffolding is possible. Regarding educational impact, it is shown that in recorded lectures, which tend to be passive, it is possible to change to active learning where learners think and explain through dialogue with virtual teachers and peers. It is also considered significant in terms of alleviating the sense of isolation in e-learning.

Limitations
As for limitations, it is stated that since the participants were limited to first-year high school students in South Korea and the subject was economics, caution is required in generalizing to other age groups or subjects. Also, as these are results from a single session, whether the effects continue with long-term use is a future task. Furthermore, there is a risk that dialogue with the LLM may deviate from the learning content, and it is stated that ingenuity is required to control this.

6. Conclusion
This study showed that dialogue with LLM-based virtual teachers and virtual peers has the potential to support metacognitive growth in self-regulated e-learning. It focuses not only on test scores but also on the accuracy of self-assessment (metacognitive monitoring). The main contributions are: (1) organizing six design requirements based on interviews, (2) implementing GPTalk using GPT-4o + RAG, and (3) demonstrating the possibility of improving metacognitive monitoring. As for future tasks, it is stated that the goal is to strengthen personalization functions according to learning history and level of understanding, and to develop it into a form that can be used long-term in more diverse learning environments.

Impressions
I felt that this study is a practical research rooted in the field, as it carefully extracted learners’ issues through preliminary interviews and conducted system design that reflects those issues. I also found the system implementation, such as the use of RAG and prompt design, very helpful. I believe it is significant that they analyzed learning support using LLMs from various aspects such as metacognitive monitoring, calibration ability, and learning experience, using multiple measurement methods rather than just learning outcomes, and examined the qualitative aspects of learning.
On the other hand, since this study was evaluated based on a relatively short 30-minute usage, I think it is necessary to verify what effects occur with long-term use. Also, I felt that the format of responding to chat in real-time while watching a video might increase the cognitive load for some learners.
I also think it is necessary to verify the accuracy of the utterances of the virtual teacher and peer roles. As mentioned in the limitations, I thought that if the content and difficulty of questions were adjusted according to the learner’s learning history and prior understanding, it could develop into a more adaptive system. I would like to apply the insights regarding staged learning support based on Self-Regulated Learning and prompt design shown in this study to my own research.