Open research questions in Intelligent Tutoring Systems and Adaptive Learning

What the literature leaves open

• (1) Longitudinal validation studies with real student populations across diverse educational contexts to validate the effectiveness of the adaptive learning algorithms and measure actual learning-outcome improvements relative to traditional platforms. (2) Implementation of federated learning using NVIDIA FLARE or Flower frameworks to enable multi-institution improvement without sharing student data, model achieving collective intelligence while preserving privacy through differential-privacy guarantees. (3) Integration of Graph Neural Networks (GraphSAGE, Graph Attention Networks) for enhanced knowledgegraph reasoning, concept-dependency modeling, and learner-similarity analysis. (4) Exploration of quantum-inspired optimization algorithms (simulated annealing using Qiskit and the Ocean SDK) for curriculum sequencing across the 10 combinatorial space of more than 10^20 possible learning paths. (5) Implementation of meta-learning (MAML) for rapid adaptation to new learners within five to ten interactions, reducing the cold-start problem. (6) Development of multimodal assessment capabilities integrating video analysis, gaze tracking, and voice analysis for richer behavioral-signal collection. (7) Expansion of causal-inference capabilities using docalculus and propensity-score matching to identify which teaching strategies causally improve outcomes rather than merely correlating with them. 13. CONCLUSION This paper presented Lumina, a next-generation, AI-powered, self-hosted learning management system that integrates multiagent orchestration, hybrid knowledge tracing, reinforcement learning, and behavior-aware personalization into a single adaptive platform. The manuscript strengthens the original contribution by formalizing the mathematical definition of the scoring and reward functions, grounding the evaluation in both synthetic and real public data, and explicitly defining a benchmark-plus-ablation protocol for future knowledgetracing comparisons. The key contributions include: (1) a closed-loop multi-agent architecture for tutoring, assessment, analytics, intervention, and governance; (2) a mathematically consistent BKT+DKT learner-modeling stack with interpretable scoring and policyoptimization equations; (3) a behavior engine that uses more than fifty passive signals and demonstrates external validity on xAPI-Edu-Data with 0.800 accuracy and 0.804 macro-F1; and (4) a benchmark-ready evaluation matrix spanning simulation, real-data validation, public-sequence datasets, and ablation analysis.

  Lumina: An Intelligent Multi-Agent Adaptive Learning Management System with Bayesian Knowledge Tracing, Deep Knowledge Tracing, and Reinforcement Learning for Personalized Education · 2026 · DOI

• The development of an adaptive student knowledge-tracing methodology within intelligent educational systems is presented in this article. The proposed system attempts to model both the complex dependencies between various forms of educational interactions and how students learn by employing heterogeneous graph constructions, graph transformer attention methods, and self-supervised contrastive learning to acquire and apply self-generated data from student interactions with digital materials. The proposed methodology implements graph attention mechanisms employing multi-head attention, enabling the acquisition of long-range conceptual connections and contextual representations of education. The experimental evaluation of benchmark educational datasets has shown that the 197 International Academic Journal of Science and Engineering, Vol. 13, No. 2, pp. 189-199. proposed framework achieves a percentage of the Knowledge State Estimation Rate is 96.8%, a Learning Path Score is 94.7%, an Interaction Stability Index is 93.1%, a Graph Representation Consistency is 92.6%, and a Recommendation Optimization is 91.4%. Comparative evaluations against previous state-of-the-art models have shown that the proposed framework performed better on all measures than the previous models. The proposed framework creates an accurate and scalable method to develop adaptive recommender systems, personalized analytics of education, and intelligent predictions of student performance within current learning environments. Future investigations could expand the proposed framework's use into multimodal adaptive learning environments that utilize textual, visual, and behavioral educational data to improve student modeling. Other possibilities of improvement for this area of research might involve applying federated self-supervised learning methods to develop privacy-preserving systems of educational intelligence and lightweight graph transformer architectures to support real-time deployment of educational intelligent systems on large-scale online learning environments. Using explainable and cognitive reasoning-based educational recommendations represents a future direction for research with the potential to improve the interpretability and effectiveness of personalized learning in the development of the next generation of intelligent education environments.

  A Graph Transformer–Based Self-Supervised Learning Framework for Modeling Student Knowledge Tracing in Adaptive Learning Systems · 2026 · DOI

• The use of AI-assisted feedback systems has steadily increased in higher education, yet there has been limited research as to which components of expert feedback can be supported by AI and which require continued human judgment.

  Human-in-the-Loop AI Feedback in Interpreter Training: An ASR-Based Platform Analysis of Instructor Annotations, Comment Functions, and System Constraints · 2026 · DOI

• regarding expected vs worst-case performance. 2 Error Classifier: Helped categorize student errors by distinguishing them into three categories: conceptual errors (failure to understand what O(n) means), structural errors (incorrectly reasoned recursive steps), or procedural errors (incorrect partition steps). 3 Progress Tracker: Monitored students' progress toward mastering recurrence reasoning, partitioning logic, and comparing computational complexity, and modified further instruction as necessary. 4 Real-time Feedback Loop: Provided instantaneous responses to students regarding their partition diagrams, choices made regarding pivots, and the linear-time properties of QuickSelect. The approach emphasized encouraging students to arrive at their own correct conclusions rather than providing them with direct answers.

  AI-Augmented Complexity Learning: Design, Automation, and Learning Impact for Conceptual Mastery in Derandomization Through Intelligent Tutoring and Real-Time Feedback · 2026 · DOI

• (16); recommendation systems and automatic LO A36; A10; A23; A41; A28; A4; labeling (7); diagnostic analysis of misconceptions (1).

  Artificial intelligence in physics education: A systematic review of content coverage, implementation models, learning impact, and pedagogical challenges · 2026 · DOI

• https://doi.org/10.1080/09500693.2010.500338 Int. 13. Asem, E.K. and Rajwa, B., Impact of combination of short lecture and group discussion on the learning of physiology by nonmajor undergraduates. Adv. Physiol. Educ., 2023, 47(1): 1–12. https://doi.org/10.1152/advan.00022.2022 14. Awwad, F., Enhancing Electronics Courses Education: Active Learning Strategies for Undergraduate Engineering Students. International Journal of Engineering Pedagogy, 2025, 15(2): 42–73. https://doi.org/10.3991/ijep.v15i2.51739 15. Margolin, J., Ba, H., Friedman, L.B., Swanlund, A., Dhillon, S. and Liu, F., Examining the impact of a play-based middle school physics program. Journal of Research on Technology in Education, 2021, 53(2): 125–139. https://doi.org/10.1080/15391523.2020.1754973 16. Hartley, K., Hayak, M. and Ko, U.H., Artificial Intelligence Supporting Independent Student Learning: An Evaluative Case Study of ChatGPT and Learning to Code. Educ. Sci. (Basel), 2024, STEM Education Volume 6, Issue 4, 539–583 565 14(2): 120. https://doi.org/10.3390/educsci14020120 17. Bitzenbauer, P., ChatGPT in physics education: A pilot study on easy-to-implement activities. Contemp. Educ. Technol., 2023, 15(3): ep430. https://doi.org/10.30935/cedtech/13176 18. Sirnoorkar, A., Zollman, D., Laverty, J.T., Magana, A.J., Rebello, N.S. and Bryan, L.A., Student and AI responses to physics problems examined through the lenses of sensemaking and mechanistic reasoning. Computers and Education: Artificial Intelligence, 2024, 7: 100318. https://doi.org/10.1016/j.caeai.2024.100318 19. Kilde-Westberg, S., Johansson, A. and Enger, J., Generative AI as a lab partner: a case study. Phys. Rev. Phys. Educ. Res., 2025, 21(2): 020119. https://doi.org/10.1103/ggy1-3kjk 20. Dahlkemper, M.N., Lahme, S.Z. and Klein, P., How do physics students evaluate artificial intelligence responses on comprehension questions A study on the perceived scientific accuracy and linguistic quality of ChatGPT. Phys. Rev. Phys. Educ. Res., 2023, 19(1): 010142. https://doi.org/10.1103/PhysRevPhysEducRes.19.010142 21. Jang, H. and Choi, H., A Double-Edged Sword: Physics Educators’ Perspectives on Utilizing ChatGPT and Its Future in Classrooms. J. Sci. Educ. Technol., 2025, 34(2): 267–283. https://doi.org/10.1007/s10956-024-10173-1 22. Durgungoz, A. and Durgungoz, F.C., Exploring effortless AI-generated gamified quizzes in an online special education module: evaluating question quality, student engagement, and its potential to identify at-risk students. Educ. Inf. Technol. (Dordr), 2025, 30(17): 25335‒25357. https://doi.org/10.1007/s10639-025-13765-5 23. Belkina, M., Daniel, S., Nikolic, S., Haque, R., Lyden, S., Neal, P., et al., Implementing generative AI (GenAI) in higher education: A systematic review of case studies.

  Artificial intelligence in physics education: A systematic review of content coverage, implementation models, learning impact, and pedagogical challenges · 2026 · DOI

• Therefore, future research should focus on adapting ICT and AI-based systems to local educational contexts, ensuring compatibility with national curricula and institutional resources.

  Adaptive AI-Supported Learning Environments in Engineering Education: Effects on Learning Outcomes, Engagement, and Digital Competence Development · 2026 · DOI

• However, future research should explore this further. When used effectively, teachers can integrate GAI into their teaching. Another point to note is that GAI chatbots produce the desired responses when users employ Python code compared to DALL-e. Furthermore, GAI chatbots struggle with representation levels, where they provide correct-looking responses with incorrect representation. This may require a more knowledgeable user. Moreover, although GAIs does well in producing texts, it personifies particles during definitions. This may lead students to view particles as living beings, creating more misconceptions. Lastly, the chatbots demonstrated the ability to aid in the conceptualization of PBL tasks by suggesting topics that can be assigned to different groups, producing drafts and rubrics, and structuring inquiry stages and preliminary simulations, among others. Nonetheless, what is crucial is that despite all these GAI abilities, the teacher would still have to demonstrate their objectivity, PCK, and CK throughout the prompting stages. Furthermore, teachers should not only rely on the responses from chatbots without verifying them with other sources. Moreover, teachers should not just use the responses from the chatbots as final drafts, but as initial drafts that would still be modified after verification with other sources and the user’s own assessment.

  Teaching and Learning Chemistry for the 21st Century Skills Through Artificial Intelligence - A Narrative Review · 2026 · DOI

• Conclusion This study demonstrates that AI-generated Shadow Podcasts can serve as an effective and well-received supplementary learning tool in higher education. In a context where teaching staff face increasing time pressures, the ability to generate weekly podcasts quickly and with minimal intervention was a key part of the pilot design. With the aim of the paper being to capture insights into whether the podcasts were used as revision tools or preparation aids for assessments and coursework, the results show that AI-generated podcasts can be of significant benefit as supplementary and supporting materials. The majority of students reported that the podcasts improved their understanding of course material, made studying more enjoyable, and helped them feel more prepared for assessments. Their concise format, conversational style, and flexibility to support multitasking were particularly valued, distinguishing them from traditional lecture recordings or slides. Crucially, students saw the podcasts not as a replacement for teaching but as a useful complement; particularly for reviewing content, revision purposes, catching up on missed material, © 2026 Journal of Perspectives in Applied Academic Practice 34 Journal of Perspectives in Applied Academic Practice | Vol 14 | Issue 1 (2026) Shadow Podcasts: Student perspectives on AI-generated audio content as a supplementary learning tool reinforcing learning in accessible ways and when preparing for assessments. This reinforces existing literature around the value of diverse learning formats and highlights a growing appetite among learners for flexible, media-rich study aids (Arkün-Kocadere & Çağlar-Özhan, 2024; Do et al., 2024; Hernandez-Lopez & Mendoza-Jimenez, 2025; Pirie & Keenan, 2025). However, students also identified clear limitations, including the artificial tone of AI voices, lack of visual support, and inconsistencies in depth or coverage. These findings point to important areas for refinement, including the need for greater emotional realism, tighter alignment with assessments, and improved accessibility features like subtitles or video integration. Overall, the findings suggest that with thoughtful implementation, generative AI tools like Shadow Podcasts can enhance blended learning environments, support varied learner needs, and serve as a vehicle for building AI literacy. Shadow Podcasts offer a scalable, efficient means of enriching the learning experience, and present a promising model for integrating generative AI into higher education in a pedagogically meaningful way. As higher education continues to adapt to the realities of digital transformation, such tools offer a promising path for innovation in content delivery and student engagement.

  Shadow Podcasts: Student Perspectives on AI-Generated Audio Content as a Supplementary Learning Tool · 2026 · DOI

• This study contributes evidence that AI-powered avatars can play a transformative role in medical education, particularly in preparing learners for complex, emotionally 100 International Journal of Advanced Corporate Learning (iJAC) iJAC | Vol. 19 No. 2 (2026) AI-Powered Avatars in Medical Education: Advancing Virtual Coaching for Clinical Readiness charged, and ethically sensitive encounters. By offering adaptive, repeatable, and unbiased practice opportunities, Conversation Mastery addresses limitations inherent in traditional lectures, role-play, and standardized patient encounters. The pilot results demonstrate that learners trained with avatars not only achieve higher success rates in conflict de-escalation but also develop greater confidence and adherence to best practices. Beyond quantitative outcomes, the qualitative feedback highlights an equally important dimension: learners valued the ability to practice “without judgment,” repeat encounters as needed, and receive structured, objective feedback. These findings resonate with the principles of deliberate practice, suggesting that avatars can provide the high-frequency, feedback-rich rehearsal that underpins skill mastery. From a broader perspective, Conversation Mastery represents more than a technological innovation; it is a pedagogical shift. Its design reflects a commitment to responsible AI use—ensuring transparency, inclusivity, accessibility, and data privacy. The system complements rather than replaces human faculty, freeing educators to concentrate on higher-order reflection and debriefing. By enabling cultural and linguistic customization, it also supports global scalability, reducing inequities in access to high-quality simulation-based education. Looking ahead, several priorities emerge for future research and development: 1. Longitudinal Impact – Studies should examine whether improvements achieved through avatar-based practice translate into sustained performance gains in clinical settings. 2. Multi-Center Validation – Expanding pilots across diverse institutions and cultural contexts will test generalizability and support global adoption. 3. Curricular Integration Models – Research should explore optimal ways to embed avatar-based training within medical and nursing curricula, from undergraduate education to residency and continuing professional development. 4. Advanced Analytics – Future iterations of the platform may incorporate predictive learning analytics, enabling early identification of learners at risk of underperformance and supporting personalized remediation. 5.

  AI-Powered Avatars in Medical Education: Advancing Virtual Coaching for Clinical Readiness · 2026 · DOI

• (1) Experiments on a larger scale: A common goal in the suggestions for future work in the reviewed studies is to expand the experiments by collecting data from a larger target group. Authors often mention this as an approach to develop more reliable and generalizable methods or systems. These improvements could help resolving drawbacks from Limitations 1 and 2. (2) Longitudinal Studies: The research presented in the reviewed studies is often evaluated by means of data collected in brief experiments in confined settings. Some studies mention that conducting a longitudinal study would significantly improve the validity of the work, which can give more realistic results than the confined experiment results mentioned in Limitation 4. (3) Increasing Modality: Lack of multimodality affects the validity and the reliability of the results, as stated in Limitation 3. Therefore, in many of the reviewed studies the authors suggest to expand their work by adding more modalities by employing additional devices or applications. This could also implicitly help with Limitations 6 and 7, since a larger variety of modalities can help overcome data errors and hardware issues Ulusoy et al. Research and Practice in Technology Enhanced Learning (2027) 22:12 Page 25 of 35 by providing more types of data. In addition, we expect that reliability and validity of the collected data increases with more data points collected through various means. (4) Exploration of Methods and Algorithms: Many studies mention that they want to expand the experiments by using different methods or algorithms. Machine learning and especially deep learning is often employed in the studies, and authors mention that using a wider variety of methods and algorithms might give better results or more precise insights. This would be a solution for Limitation 5, which discusses the issues emerging from the lack of algorithm and parameter variety.

  Learning analytics with multimodal data through the lens of AI in education: A systematic literature review · 2026 · DOI

• In previous work (Samuelsen et al., 2021), we identified challenges of using xAPI for learning context description through a systematic analysis of data originating from 1) interviews with the stakeholders within the AVT project, and 2) inspection of the xAPI and xAPI profile specifications (Advanced Distributed Learning, 2018a; Advanced Distributed Learning, 2021). Thereafter, we recommended solutions to support interoperability and data integration, with emphasis on descriptions of xAPI context, while also providing some recommendations that relate more generally to the xAPI framework (i.e., recommendations for data typing and validation, and documentation). For the list of recommendations, please refer to Samuelsen et al. (2021).

  Implementing enriched context descriptions for efficient scaling of learning analytics · 2026 · DOI

• This paper addressed two research questions related to the implementation and evaluation of recommendations for enhancing xAPI context descriptions and expressibility that were identified in previous research (Samuelsen et al., 2021). In response to RQ1—how can the recommendations for enhancing xAPI context descriptions and expressibility be implemented through a technical solution?—the recommendations for enhancing xAPI context descriptions and expressibility were implemented as part of a technical solution through the creation of a new xAPI profile for the K-12 adaptive learning domain, where an accompanying xAPI example statement illustrates data description according to the profile. The technical solution incorporates various aspects, including a unified and hierarchical context model that is exemplified with categorization of context according to the adaptive learning domain, emphasizing measures for data typing and validation. In response to RQ2, how does the technical solution support technical experts in describing learning activity data from multiple data sources for LA data integration?, technical experts participating in user testing of the technical solution expressed that the solution supports them in describing learning activity data for LA data integration consistently across sources. As such, the user test participants generally indicated approval of the unified and hierarchical structuring of context, as represented through context dimensions and properties, emphasizing that it can add to aspects related to expressibility. The participants, however, indicated pragmatic views that consistent data descriptions can alternatively be achieved through other means such as post hoc data processing. Furthermore, the participants indicated general support for data typing and validation measures to support data integration, although cautioning that these need to follow realworld constraints. The participants also indicated approval for the use of hosted metadata to promote consistent data descriptions. Related to the usability criteria assessed for the technical solution, the user testing generally indicates that the usefulness and effectiveness criteria are met. Additionally, the participants indicate satisfaction regarding the solution at a general level, while also pointing out potential for improvement at the implementable level. The findings of this study provide evidence that the technical solution is feasible and effective in supporting xAPI data integration and interoperability through promoting expressibility, focusing in particular on the K-12 adaptivity case. The value contribution of this study is improved xAPI data integration through the promotion of consistent data Samuelsen et al.

  Implementing enriched context descriptions for efficient scaling of learning analytics · 2026 · DOI

• Some limitations of this study must be acknowledged. For example, learners with fewer than 50 interactions were excluded, limiting the applicability of the proposed approach in sparse data settings involving many new or infrequent users.

  A Bandit-Based Approach to Educational Recommender Systems: Contextual Thompson Sampling for Learner Skill Gain Optimization · 2026 · DOI

• One limitation of this study is that the chatbot required us to design pre-planned responses based on intents assigned by the AI. Finally, this study was limited by the timing of the experience in relation to other learning experiences that PSTs had in the methods class, their field experiences, and in the teacher education program more broadly.

  Cultivating responsive teaching with AI: exploring preservice teachers’ questioning patterns with student-emulating agents · 2026 · DOI

• Instructor governance layer Gate state, policy version, required evidence, specificity check, bypass attempt, override. Allowed prompt type, hint level, retrieval source, response, guard result, revision link. Assumptions, plans, tests, explanations, timestamps, versions, peer comments. Evidence-linked depth profile, confidence, missing-evidence flags, rule/model version. Role participation, critique, uptake, revision, and group reflection. Gap type, resource, source rationale, acceptance, and completion. Task configuration, audit sampling, intervention, override, follow-up, and review trail. D.

  ThinkDeeper Web Coach: A Literature-Informed Design Specification for Process-Visible Reasoning in Problem-Solving-Intensive Information Technology Courses · 2026 · DOI

• To integrate the mechanisms, this article introduces the Process-Visible Learning Design Framework (PVLDF). The framework links evidence elicitation, scaffolded inquiry, traceable revision, interpretive analytics, gap-responsive Figure 1. Literature-informed Design and Development Research Procedure Used in the Study. DOI: 10.34148/teknika.v15i2.1504 TEKNIKA, Volume 15(2), July 2026, pp. 370-379 ISSN 2549-8037, EISSN 2549-8045 372 Arrasyid, R. et al.: ThinkDeeper Web Coach: A Literature-Informed Design Specification for Process-Visible Reasoning in Problem-Solving-Intensive Information Technology Courses B. Data Sources and Screening The empirical foundation was constructed from academic records indexed in Semantic Scholar and OpenAlex. The search export used for analysis was generated on 24 December 2025 and covered publications from 2023 through 2025. The exact natural-language semantic query was: "ThinkDeeper Web Coach: Discourse-Aware Learning Analytics + Learning Circle Orchestration + Resource Recommendation to Increase Students' Depth of Thinking in IT Courses." No explicit language filter was applied. Journal articles, conference papers, and preprints were eligible when they provided empirical or design-oriented educational through evidence. Duplicate candidates were normalized title, DOI, author, and venue metadata and were not treated as independent evidence during synthesis. flagged Figure 2. Evidence Identification and Screening Process Used to Derive the Artifact Requirements. synthesis. Screening The initial corpus contained 500 records. Abstract screening excluded 370 records and retained 130 reports for retrieval. Full text could not be obtained for 77 reports; 53 full texts were assessed against the six criteria, two were excluded because the target population fell outside the intended IT/computing boundary, and 51 studies entered extraction and requirement target population, technology sophistication, cognitive or selfregulatory outcomes, empirical or design evidence, formal educational setting, and relevance to the artifact mechanism. Figure 2 and Table 2 report the reproducible flow and decision rules. Because the corpus is heterogeneous and includes adjacent disciplinary contexts, identify mechanisms and boundary conditions rather than to estimate pooled causal effects. considered is used to it judgments and Criterion-level yes/maybe/no their rationales were recorded in structured worksheets. Uncertain abstract-stage records were retained when the holistic judgment favored inclusion and were resolved during full-text assessment. A structured model-assisted first pass was followed by consistency review; the process was not a blinded multi-reviewer screening, and no inter-rater agreement coefficient was calculated. Reports that could not be retrieved were recorded separately and were not counted as full-text eligibility exclusions. Table 2.

  ThinkDeeper Web Coach: A Literature-Informed Design Specification for Process-Visible Reasoning in Problem-Solving-Intensive Information Technology Courses · 2026 · DOI

• Indeed, a recent study reported in Harvard Business Review concluded that the mixed results from employee use of AI to increase productivity was correlated to those who “monitor” their thinking vs.

  Asked & Answered · 2026 · DOI

• However, most LLM based learning systems are designed for either single users or symmetric collaboration, leaving parent child tutoring with distinct instructional roles underexplored.

  ParaTutor: LLM Mediated Parent Child Tutoring through Role Separated Scaffolding Interface in Real Time · 2026

• Building on the conceptual analysis, bibliometric findings, and empirical data presented in this study, the following actionable recommendations are proposed for key education stake- holders to support the responsible integration of AI into teaching and teacher education: For Teacher Educators: Integrate the AIA-PCEK framework into both pre-service and in-service teacher educa- tion curricula. Particular emphasis should be placed on developing competencies in: ● AI-agent literacy, ● Ethical and responsible AI use, and. ● Adaptive instructional design, ● in alignment with the five domains of the UNESCO (2024) AI Competency Framework for Teachers. Professional learning should encourage reflective practice, critical engage- ment with AI, and human–AI co-design approaches. For Policymakers: Develop and implement national and institutional strategies that support: ● The ethical regulation of AI in education, ● Mandatory professional development frameworks, and. ● Equitable access to AI-enhanced learning tools. ● Policies should be aligned with international standards (e.g., UNESCO, OECD) and promote inclusion, data transparency, and teacher autonomy in AI-rich environments. For Educational Technologists and Developers: Engage in collaborative design processes with educators to build AI systems that are: ● Pedagogically grounded, ● Explainable and fair, and. ● Inclusive by design. ● Ensure that principles such as user control, accountability, and non-discrimination are embedded in all AI-driven educational technologies. Design features should also enable teachers to moderate, adapt, and contextualize AI-generated outputs in real time. For Researchers: Undertake rigorous empirical studies using mixed-method designs to evaluate: ● The effectiveness and usability of AIA-PCEK in diverse settings; ● Its alignment with teacher competencies, as defined by UNESCO and other global frameworks; 1 3A. Mimoudi, A. Bouabid ● Its impact on teacher agency, student autonomy, and ethical classroom practices over time. ● Cross-national studies and longitudinal designs will be particularly valuable in testing the scalability, adaptability, and policy relevance of AIA-PCEK across different cultural and institutional contexts. Acknowledgements The author gratefully acknowledges the support of the Erasmus+ VOLCANIC Project and the Institute of Education Sciences at Mohammed VI Polytechnic University (UM6P). Sincere thanks are extended to the participating teachers and facilitators involved in the training activities that informed this study. Author contributions A.M. conceived the study, designed the research framework, conducted the biblio- metric analysis, coordinated the empirical data collection from the VOLCANIC project, developed the AIA-PCEK framework, and wrote the original draft of the manuscript. A.M. also prepared all figures and tables.A.B. contributed to the conceptual refinement of the framework, provided critical input on the policy alignment with the UNESCO AI Competency Framework for Teachers, and participated in the review and revision of the manuscript.All authors reviewed and approved the final version of the manuscript. Funding The author received no financial support for the research, authorship, or publication of this article. Data availability No datasets were generated or analysed during the current study.

  AIA-PCEK: a new framework for responsible AI integration in teacher education · 2026 · DOI

• Addressing data privacy, algorithmic bias, transparency, and accountability in AI-mediation education Designing human-centered AI-enhanced activities; supporting inquiry, collaboration, and critical…

  AIA-PCEK: a new framework for responsible AI integration in teacher education · 2026 · DOI

• To explore the effects of our chatbot-based learning journaling system, we conducted an experimental field study with a three-week usage period. This has several limita- tions. While the study was conducted over multiple weeks, it is still difficult to derive conclusions regarding the long-term effects of the proposed design principles. In par- ticular, the declining rate of journal entry creation after course completion suggests that longer multi-phase interventions are needed to fully understand how to best uti- lize the proposed design principles. In addition, while the application usage (except for the onboarding session) was not compensated, and this fact was communicated to the participants before they started the study, the participants might still have felt required to use the application because of the study setting and the fact that they were part of a panel for study participation. Future work could extend on this and the previous point by providing a version of the app after the study ends to monitor users’ long-term behavior. Another potential issue is the environment and timing of the study: Because the sur- veys were provided at different times of the semester, there might be confounding effects due to seasonal and semester effects influencing the well-being of students (Lukmanji et al., 2020; Pitt et al., 2018). Because of the interconnectedness between SRL and well- being, seasonal and semester effects might influence the reported SRL (Boekaerts, 2011). In addition, while the free-will approach to using the assistant and summary fea- tures should improve external validity, it comes at the risk of internal validity. On the one hand, the assistant was not used enough to impact most participants. This could either result from the assistant’s lack of perceived value or because the assistant was not as tightly integrated as the summaries or the course. On the other hand, from the per- spective of personalized support or adaptive learning, it seems to make sense that the AI assistant was useful for a specific group of participants for promoting motivation or maintaining engagement and that other participants did not work with the AI assistant Scheu et al. International Journal of Educational Technology in Higher Education (2026) 23:15 Page 20 of 24 at all or applied it only at the beginning of their reflection process. In general, a challenge for future research activities on adaptive learning will be to analyze which type of sup- port is appropriate for particular student subgroups at specific time points. Neverthe- less, in the present study a self-selection bias regarding the engagement of participants that used the assistant cannot be ruled out. Lastly, our implementation is just one instantiation of the design principles we laid out. The observed effects might not always result from the general concept of the intervention but from issues in our implementation. Future work could investigate the existing features in isolation or require the participants to use them to gather more meaningful data on their effects, provide alternative implementations, and explore new tools. Another research avenue would be to investigate the effects of summarization on the reflection process. In this study we included them for all participants to improve the value of the chatbot and did not isolate their effects. A further limitation concerns the operationalization of engagement. Our primary behavioral metric was response length in characters, complemented by journal-entry timing and notification reliance. This does not capture cognitive engagement, reflective depth, or the quality of the produced journal entries. Future work should therefore com- bine log-based measures with qualitative coding of reflection quality in addition to the measures presented here.

  Designing a mobile chatbot-based learning journaling system for intrinsic motivation and engagement · 2026 · DOI

• There are two significant limitations of the study: First the research team was unable to implement the lessons in a school setting, which may have dampened students’ enjoy- ment, as multiple students mentioned that the lessons were similar to school. The setting (i.e., public spaces) may have affected students’ responses to the review questions, as stu- dents’ learning ability in formal settings does not always correspond to their learning ability in informal settings (Griffin, 1994). Second, the low sample size prevents gen- eralizations from this study to other ITSs or other design elements within iSTART-Early (e.g., students’ enjoyment of the meta-game in iSTART-Early). To address these two limitations, the next step in this research and development project will be to assess the functionality and usability of the complete iSTART-Early system for both teachers and students in classrooms. Future work will continue the cur- rent studies’ iterative design to continuously implement teachers’ and students’ feedback into the system to enhance learning experiences and outcomes for the end users. In addition, iSTART-Early will be made freely available upon completion of the development. Both a larger, classroom- based study and individual children’s use of the platform will provide greater insight into the usability and efficacy of iSTART-Early. Classroom-based research on iSTART- Early will also afford teachers and researchers the ability to assess how and for whom iSTART-Early can be used to teach reading strategies.

  Engineering the Design of iSTART-Early: Adapting an Intelligent Tutoring System for 3rd, 4th, and 5th Grade Reading · 2026 · DOI

• This paper presents a detailed design and evaluation framework, but it does not report large-scale experimental results. Therefore, claims about effectiveness should be treated as proposed or expected rather than proven. A full version of the study would require user testing over several weeks or months. Another limitation is that student behavior is highly individual. A planning algorithm that works for one learner may not work for another. Some students prefer fixed schedules, while others prefer flexible task lists. The application must therefore support personalization rather than enforcing one planning style.

  Studyflow: The Smart Study Planner · 2026 · DOI

• virtual agents, and collaborative platforms to enable personalized, scalable, and collaborative mentorship. Proposed an AI-augmented realitybased virtual mentorship system for learning. Methods student include personalized support, contextual learning pathways, real-time feedback, gamification, AI avatars, NLP, and AR technologies to improve understanding of AI concepts. stop using chatbots and examine factors related long-term user to loyalty and retention. Enhance AI personalization, improve multimodal interactions, create hybrid AI-human mentorship models, ensure ethical data cultural practices, and support experiential learning. Policy frameworks and crosscultural adaptation recommended.

  CSE Mentor: an ML-Driven Virtual Mentor for Department · 2026 · DOI

Most-cited papers in Intelligent Tutoring Systems and Adaptive Learning

• What Is the Impact of ChatGPT on Education? A Rapid Review of the Literature · Education Sciences · 2023 · 1,415 citations
• ChatGPT: Bullshit spewer or the end of traditional assessments in higher education? · Journal of Applied Learning & Teaching · 2023 · 931 citations
• Role of AI chatbots in education: systematic literature review · International Journal of Educational Technology in Higher Education · 2023 · 745 citations
• Interacting with educational chatbots: A systematic review · Education and Information Technologies · 2022 · 620 citations
• The impact of artificial intelligence on learner–instructor interaction in online learning · International Journal of Educational Technology in Higher Education · 2021 · 556 citations
• The promise and challenges of generative AI in education · Behaviour and Information Technology · 2024 · 343 citations
• Promises and challenges of generative artificial intelligence for human learning · Nature Human Behaviour · 2024 · 328 citations
• Exploring the potential of artificial intelligence tools in educational measurement and assessment · Eurasia Journal of Mathematics Science and Technology Education · 2023 · 314 citations
• Leveraging AI in E-Learning: Personalized Learning and Adaptive Assessment through Cognitive Neuropsychology—A Systematic Analysis · Electronics · 2024 · 247 citations
• Benchmarking Retrieval-Augmented Generation for Medicine · 2024 · 210 citations

Most recent work

• Neuro-symbolic synergy in education: a survey of LLM-knowledge graph integration for explainable reasoning and emotion-aware student support · Smart Learning Environments · 2026
• Creating Assessments Using Artificial Intelligence: An Introduction to the TEACH-SAN-TA Model · Journal of Physical Education Recreation & Dance · 2026
• AI self-efficacy and knowledge graph-integrated generative AI feedback in higher education · The Internet and Higher Education · 2026
• AI-driven adaptive stealth assessment for socially regulated learning in collaborative environments · Journal of Research on Technology in Education · 2026
• A systematic review of intelligent and robot tutoring systems: evolution, pedagogical design, and AI-driven classification · Smart Learning Environments · 2026
• An integrated framework for outcome based education and AI supported blended learning in curriculum redesign and intelligent training management · Discover Computing · 2026
• Using Artificial Intelligence in Test Construction: A Practical Guide · Psicothema · 2026
• Innovative GPT-Powered Adaptive Assessment for Mobile Language Learning: A Path to Real-Time Personalization and Enhanced Engagement · International Journal of Linguistics Literature & Translation · 2026
• The 4C Framework: Integrating AI in Creating and Solving Case Studies · Journal of Advertising Education · 2026
• Rex to the rescue: Evaluating the effectiveness of real-time AI-driven stealth assessment and support for mathematical modeling in virtual science investigations · Journal of Research on Technology in Education · 2026

Related topics in Computer Science