Adaptive learning today is a leading edtech trend, aiming at responding to the challenges of mass education with a personalized approach. Research on adaptive learning has been broadly conducted in a number of academic centers across Europe. Our literature revealed analysis of adaptive learning contexts at German Universities (Hemmler, 2022), studies of adaptive textbooks in the Netherlands (Sosnovsky, 2023) and research on the use of artificial intelligence in the area in Cyprus (Zacharioudakis, 2024). Artificial intelligence, neural networks and LLMs are becoming important assets in developing the capabilities for adaptive learning, in particular in creating resources providing multiple learning paths for learners. Still an obvious lack of evidence extracted from actual adaptive learning platforms, that could be used as proof of research hypotheses is observed. To overcome this drawback we propose our paper, presenting results of 01Math adaptive learning system practical evaluation in Russian middle school science classroom settings in 2023 - 2024. The system, developed in September, 2023 currently contains adaptive learning resources for teaching science to 5 and 6 grades students within the framework  of personalized approach. Each of 120 units contains a short test fulfilled by students. Depending on their performance students are either directed to more advanced units or advised to cover the material again in a more detailed and explicit manner. 

Our paper aims at the research question: “how can the power of large language models be applied to enhance the effects of adaptive learning in science classroom settings”? As practitioners, evaluating our adaptive learning systems in classrooms we see how difficult it is to provide the appropriate amount of teaching resources and testing materials. Our first experiment with large language models show that they can be effectively used in creating the necessary number of artificially generated tasks. Still the appropriate methods of building the most efficient prompts and evaluating the resulting materials are yet to be found. We see finding answers to these questions as the objective of our discussion.

Principles of instructional design and concept learning are used as a theoretical framework of our research (Merrill, 1997). According to them, learning can be represented and modelled as an interplay of two sets of values, the subject domain and the students’ domain. In our approach the subject domain is understood as a network of related concepts, each of which is an essence of understanding a topic within a taught discipline (Tennyson, 1986). The use of natural language processing (NLP) tools and procedures makes it possible to define prerequisites and post requisites in the concept network. At the preliminary stage of our research we built a concept network for the middle school science course, which became the basis for the subject domain construction in our adaptive learning system. At the production stage all 200 content units were supplemented with tests, triggering students’ future paths. Platform records, reflecting students’ actions on the platform were collected thus providing necessary data for analysing the student’s domain variables.

Methodology, used for task generation, which we applied in our research used benchmarking of LLM generated adaptive learning tests across human expert questions writing (Kıyak, 2024) . A number of possible applications of LLMs in adaptive learning has already been described in the literature on the topic: -Facilitating platform notifications for students -Summarizing search results on a topic -Generating examples for a concept presentation -Cases and quizzes Generation of unit tests in our case was performed with Chat GPT-4 from OpenAI. For evaluating LLM generated adaptive learning resources three groups of metrics and benchmarks are used. Concurrence metrics are applied when a simple method of calculating target words in source text and prompt generated artefacts can be applicable. Neural network metrics use embedding vectors (BERT, Yisi) for comparing model and generated text similarities (Tran, 2023). For evaluating generated multiple choice questions it is essential to measure distractors quality. In our research we used for this purpose their credibility and consistency. Performance assessment in this case was fulfilled with a non-reference metric. Unreliable distractors were identified by human experts but reliability of students’ responses was inspected with a LLM.

Our research has led to several observations regarding the use of adaptive learning in teaching science at middle schools: -Both teachers and students are willing to use digital adaptive learning systems in their educational practice. Students are particularly eager to study online, as they use their mobile devices for most of their communications. -There is a significant interest in adapting the teaching content in the textbooks publishers for the use on digital platforms. Still there is a significant lack of methodology of such transformation. Most textbook content is static and linear. Adapting it to digital platforms will require full-fledged content restructuring. -Teachers experience a serious need in teaching guidelines for work in adaptive learning environments. Further research on adaptive learning as a means of transforming school textbooks into personalized digital environments, can become a seminal point in finding innovative tracks and strategies that will become crucial in shaping future European education. This project will include the following steps: -Conduct a study of best practices of adaptive learning and aligning printing textbooks content with digital platforms; -Develop a methodology of building knowledge maps for school subjects content domains based on artificial intelligence tools; -Build an instructional design strategy for integrating adaptive learning platforms into classrooms. -Develop professional criteria for educators, migrating from printed textbooks to adaptive learning systems. -Develop a teachers’ toolkit for evaluating adaptive learning systems

Chaudhri, V. K., Cheng, B., Overtholtzer, A., Roschelle, J., Spaulding, A., Clark, P., ... & Gunning, D. (2013). Inquire biology: A textbook that answers questions. AI Magazine, 34(3), 55-72. Kıyak, Yavuz Selim & Coşkun, Özlem & Budakoğlu, Işıl & Uluoglu, Canan. (2024). ChatGPT for generating multiple-choice questions: Evidence on the use of artificial intelligence in automatic item generation for a rational pharmacotherapy exam. European Journal of Clinical Pharmacology. 80. 1-7. 10.1007/s00228-024-03649-x. Hemmler, Y. M., & Ifenthaler, D. (2022, July). Indicators of the learning context for supporting personalized and adaptive learning environments. In 2022 international Conference on advanced learning technologies (ICALT) (pp. 61-65). IEEE. Merrill, D. (1997). Instructional Transaction Theory: An Instructional Design Model Based on Knowledges Objects. In R. D. Tennyson, F. Schott, N. Seel, & S. Dijkstra (Eds.), Instructional design: International Perspectives, Vol. I: Theory and research (pp. 215-241). Mahwah, NJ: Erlbaum. Sosnovsky, S., Brusilovsky, P., & Lan, A. (2023, June). Intelligent Textbooks: The Fifth International Workshop. In International Conference on Artificial Intelligence in Education (pp. 97-102). Cham: Springer Nature Switzerland. Tennyson, R. D., & Cocchiarella, M. J. (1986). An empirically based instructional design theory for teaching concepts. Review of Educational Research, 56(1), 40-71. Tennyson, R. D. (2010). Historical reflection on learning theories and instructional design. Contemporary educational technology, 1(1), 1-16. Tran, Andrew & Angelikas, Kenneth & Rama, Egi & Okechukwu, Chiku & Smith, David & Macneil, Stephen. (2023). Generating Multiple Choice Questions for Computing Courses Using Large Language Models. 10.1109/FIE58773.2023.10342898. Zacharioudakis, E., Romankevich, V. A., Zacharioudakis, S., & Tsovilis, G. (2024). ADAPTIVE LEARNING IN EUROPEAN INTEGRATION: INNOVATIONS AND CHALLENGES. In The Tenth international conference on adaptive learning. ATL-2024, 13.