Educational technologies have evolved significantly over recent decades, with AI representing the latest frontier in this progression. Current applications range from adaptive learning platforms that personalize content delivery to automated systems that provide immediate feedback. The challenge lies in developing AI educational technologies that enhance human capabilities while respecting the autonomy and agency of learners and educators. My research aims to design, implement, and evaluate human-centered AI systems in educational contexts. By prioritizing human needs and values throughout the development process, my work seeks to advance the understanding of how AI can enhance educational practices while preserving the learner's independence and the educator's role.
In the last decades, the advancements in Artificial Intelligence (AI) have enabled the development of
Conversational Agents (CAs) (e.g., ChatGPT [
However, while these tools increase eficiency and accessibility, over-reliance on such tools can
weaken learners’ critical thinking, problem-solving abilities, and overall learning experience. By
delivering directly complete solutions, CAs hinder students’ opportunity to work through problems
independently and develop their analytical skills [
Human-Centered Artificial Intelligence (HCAI) [
The application of HCAI principles in educational contexts presents both significant opportunities and challenges. In educational settings, the primary goal is not to automate tasks or provide quick answers, but to foster learning, critical thinking, and skill development. Therefore, HCAI in education can be applied to design systems to support the learning process while maintaining learner agency and promoting deep understanding.
The HCAI principles applied in the education setting are summarized in Table 1. For decades,
researchers have concentrated on developing computer systems that emulate a human tutor, guiding
learners through the learning process. The aim is to provide one-to-one personalized tutoring in
contexts where one-to-many instructions from a single teacher is not enough (e.g., traditional classroom
lectures). Intelligent Tutoring Systems (ITSs) aim to tutor learners by providing personalized feedback,
automatic assessment and by answer questions, without requiring the intervention of a human teacher,
improving the quality of education [
While significant research is dedicated to support learners, it is essential to recognize the central role of educators in the education setting. AI systems can play a valuable role in supporting teachers,
CEUR Workshop
ISSN1613-0073
easing their workload (e.g., expert decision making, tools for automated exercise and quizzes generation)
[
Finally, the design of educational systems should be aligned with real user needs [14]. AI systems are more intuitive, acceptable and efective when students and teachers are actively considered in the design process [15].
Research Subtheme Supporting Literature Conclusions Learning Analytics Tools and Personalized Rec- Adapt to the individual needs of students ommendation Systems by developing intelligent tutoring systems that help them master transferable learning skills, thereby enhancing their self-directed learning abilities.
Intelligent Systems Supporting Smart Educa- Explore the development of intelligent systion Development tems to support personalized learning and smart education to improve the quality of education.
Intelligent Tutoring Systems and Expert Enhance the interaction and collaboration Decision-Making Chatbots eficiency between teachers and students through intelligent tutoring systems and expert decision-based chatbots to improve teaching efectiveness.
Human-Centered Smart Education Models Emphasize the human-centric and nurturing aspects of teacher-student interactions, advocating for the retention of the core role of human teachers in education, and creating a new model of intelligent education for teacher-student interaction.
Emphasis on User Needs and Experience The redesign of virtual assistant design and user interfaces increases the applicability and user acceptance of the technology, reflecting the principle of human-centric design.
My dissertation work aims to exploit AI methods to enhance both student and teacher work. The objective is to support the student learning process, analyzing existing approaches’ opportunities and limitations, and develop, test and evaluate new techniques and tools to support both students and teachers in the learning journey.
An increasing number of scholars are exploring how to support learners in diferent subjects. Several works have tried to identify the best learning path to master a given math concept employing a Reinforcement Learning (RL) algorithm. Liu et al. [16] have developed a system that takes into consideration both which concepts the learner has already mastered and their releationship with several other concepts yet to be mastered to identify the sequence of exercises to suggest to the learner. Li et al. [17] have exploited a Hierarchical RL algorithm to re-adapt the sequence of exercises based on the learner’s actual knowledge and the goal to achieve. More recent works exploit Large Language Models (LLMs) to co-create stories with the learner to learn math language using voice interaction [18] and visual representation of the scenes [19].
In other works, LLMs are used to provide timely feedback and answer to learners’ questions. Kazemitabaar et al. [20] developed CodeAid, an LLM-based assistant designed to support programming students without revealing full solutions but providing help, pseudo-code, and code explanations.
Hou et al. have developed CodeTailor [21] which creates Pearson puzzles by employing an LLM. This system provides personalized help to students while encouraging the cognitive engagements.
Choi et al. [22] conducted a design workshop with educational experts and educators on the potential usage of LLMs to transform a monologue lecture script into pedagogically meaningful dialogue. Subsequently, they developed VIVID, a LLM-based system that allows co-creation of pedagogical dialogues. ReadingQuizMaker [12] supports educators to easily create high-quality multiple choices and open-ended questions by employing a (Natural Language Processing) NLP-based process starting from a text.
These works illustrate how, when intentionally implemented, AI can be leveraged to develop systems that can engage learners in mastering concepts and support educators to enhance their teaching processes.
My objective is to study these approaches and explore how to efectively integrate AI into educational systems to support meaningful learning, foster critical thinking, and enhance both student and teacher experiences.
The overall objective of my research is to explore how AI can both support learners to reach their learning goals and give teachers practical support. This research will be conducted aiming to answer the following research questions: • RQ1: How to design AI-based systems to efectively support student and teachers in everyday learning activities? – RQ1.1: What kind of interaction between learners and AI promote engagement, reflection, and active participation without hindering skills’ development? – RQ1.2 How can AI-based systems be designed to empower teachers? • RQ2: What key features and best practices can be identified in these systems to inform the design of AI-based educational tools?
This research will proceed through three phases to design educational AI-based systems that support learners and teachers, while keeping them central to the process. To answer the research questions outlined in Section 3, I plan to: 1. Asses the current status of AI-based learning. I am conducting a review of the existing literature to analyze the strenghts and the limitations of the current approaches, with particular attention to how they can help both students and teachers (RQ1). 2. Design, implement, and evaluate AI-based systems that can empower teachers (RQ1.2) and support learners (RQ1.1). To this end, it is essential to preserve their role and agency throughout the whole process. 3. Guideline derivation. Starting from the designed systems, I will extract the guidelines for the design of a HCAI educational system that can be used to augment and enhance both educators and learners’ capabilities, while increasing engagement and pleasure to learn (RQ2).
During my first year, I reviewed the literature to build the foundation steps to develop a system that can help students in their learning journey.
Existing math ITSs often employ RL algorithms to suggest sequences of exercises based on the learner’s ability to correctly and completely solve previous exercises. However, these systems are not able to capture when the student struggles to solve a specific intermediate step. This limitation is due to the fact that existing educational math datasets they are trained on do not include intermediate steps interactions.
To build a step-aware adaptive ITS, it is necessary to decompose exercises into detailed step-by-step solutions, thus enabling the training of a RL algorithm employing simulated student that interact with the system.
In my work, I leveraged GPT-o3-mini1 to break-down the solution of exercises from the Junyi educational dataset [23] into intermediate steps. If the intermediate step was still too complex, additional sub-steps were generated. The step-by-step decomposition provided by GPT-o3-mini was validated by comparing the generated exercise solution with the actual one, ultimately retaining only correct decompositions. Subsequently, I analyzed whether the generated steps and sub-steps efectively guided learners toward the correct solution.
To evaluate this, I used three models of varying sizes from the Llama family [24, 25]. Each model was ifrst asked to solve an exercise directly. If it failed, it was progressively provided with the generated intermediate steps and prompted to solve both the current step and the full exercise. If the model struggled with a specific intermediate step and sub-steps had been generated, these were also provided, and the model was asked to solve the sub-step, the corresponding step, and the complete exercise. The process is outlined in Fig.1.
Results show that the models were able to solve 42% more exercises when guided by intermediate steps, compared to when prompted to directly solve the whole exercises. These findings proved the usefulness of the generated step-by-step solutions to guide the students - in this case, simulated by models from the Llama family - toward the solution of math problems.
Currently, I am starting my second year of the National Ph.D. program in Artificial Intelligence at Politecnico di Torino under the supervision of professor Luigi De Russis.
During my first year, I have explored the existing literature studying how researchers have integrated AI in educational systems to support both students and teachers.
During my second year, I will continue developing the previously mentioned step-aware ITS. Additionally, I plan to design multimodal systems that enable stakeholders to interact with the technology through various input and output modes (e.g., voice, text). I also intend to conduct user testing with interested participants to evaluate the system’s efectiveness and usability.
In the third year, I aim to extract key design principles from my findings and define a set of guidelines for developing AI-based educational systems that can enhance and augment learners and educators capabilities while keeping their autonomy and agency central to the design process.
In conclusion, I expect my work to contribute to more responsible and efective uses of AI in education, enhancing student engagement, supporting teachers, and promoting systems that align with human-centered values.
This work was supported by the Cineca consortium.
During the preparation of this work, the author used ChatGPT in order to: grammar and spelling check, paraphrase and reword. After using this tool, the author reviewed and edited the content as needed and takes full responsibility for the publication’s content. @ Scale, L@S ’24, Association for Computing Machinery, New York, NY, USA, 2024, p. 305–309.
URL: https://doi.org/10.1145/3657604.3664694. doi:10.1145/3657604.3664694. [12] X. Lu, S. Fan, J. Houghton, L. Wang, X. Wang, Readingquizmaker: A human-nlp collaborative system that supports instructors to design high-quality reading quiz questions, in: Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, CHI ’23, Association for Computing Machinery, New York, NY, USA, 2023. URL: https://doi.org/10.1145/3544548.3580957. doi:10.1145/3544548.3580957. [13] I. Celik, M. Dindar, H. Muukkonen, S. Järvelä, The promises and challenges of artificial intelligence for teachers: a systematic review of research, TechTrends 66 (2022) 616–630. URL: https://doi.org/ 10.1007/s11528-022-00715-y. doi:10.1007/s11528-022-00715-y. [14] R. Alfredo, V. Echeverria, Y. Jin, L. Yan, Z. Swiecki, D. Gašević, R. Martinez-Maldonado, Humancentred learning analytics and ai in education: A systematic literature review, Computers and Education: Artificial Intelligence 6 (2024) 100215. URL: https://doi.org/10.1016/j.caeai.2024.100215. doi:10.1016/j.caeai.2024.100215. [15] X. Chen, Z. Hu, Y. Jing, C. Wang, The development and alleviation of humanistic artificial intelligence in the digital age education: A content analysis research, in: Proceedings of the 2024 8th International Conference on Advances in Artificial Intelligence, ICAAI ’24, Association for Computing Machinery, New York, NY, USA, 2025, p. 206–211. URL: https://doi.org/10.1145/ 3704137.3704189. doi:10.1145/3704137.3704189. [16] Q. Liu, S. Tong, C. Liu, H. Zhao, E. Chen, H. Ma, S. Wang, Exploiting cognitive structure for adaptive learning, in: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, KDD ’19, Association for Computing Machinery, New York, NY, USA, 2019, p. 627–635. URL: https://doi.org/10.1145/3292500.3330922. doi:10.1145/3292500.3330922. [17] Q. Li, W. Xia, L. Yin, J. Shen, R. Rui, W. Zhang, X. Chen, R. Tang, Y. Yu, Graph enhanced hierarchical reinforcement learning for goal-oriented learning path recommendation, in: Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, CIKM ’23, Association for Computing Machinery, New York, NY, USA, 2023, p. 1318–1327. URL: https: //doi.org/10.1145/3583780.3614897. doi:10.1145/3583780.3614897. [18] C. Zhang, X. Liu, K. Ziska, S. Jeon, C.-L. Yu, Y. Xu, Mathemyths: Leveraging large language models to teach mathematical language through child-ai co-creative storytelling, in: Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems, CHI ’24, Association for Computing Machinery, New York, NY, USA, 2024. URL: https://doi.org/10.1145/3613904.3642647. doi:10.1145/3613904.3642647. [19] H. Li, R. Guo, C. Li, W. Xing, Automated quality assessment of multimodal mathematical stories generated by generative artificial intelligence, in: Proceedings of the Eleventh ACM Conference on Learning @ Scale, L@S ’24, Association for Computing Machinery, New York, NY, USA, 2024, p. 110–121. URL: https://doi.org/10.1145/3657604.3662029. doi:10.1145/3657604.3662029. [20] M. Kazemitabaar, R. Ye, X. Wang, A. Z. Henley, P. Denny, M. Craig, T. Grossman, Codeaid: Evaluating a classroom deployment of an llm-based programming assistant that balances student and educator needs, in: Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems, CHI ’24, Association for Computing Machinery, New York, NY, USA, 2024. URL: https: //doi.org/10.1145/3613904.3642773. doi:10.1145/3613904.3642773. [21] X. Hou, Z. Wu, X. Wang, B. J. Ericson, Codetailor: Llm-powered personalized parsons puzzles for engaging support while learning programming, in: Proceedings of the Eleventh ACM Conference on Learning @ Scale, L@S ’24, Association for Computing Machinery, New York, NY, USA, 2024, p. 51–62. URL: https://doi.org/10.1145/3657604.3662032. doi:10.1145/3657604.3662032. [22] S. Choi, H. Lee, Y. Lee, J. Kim, Vivid: Human-ai collaborative authoring of vicarious dialogues from lecture videos, in: Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems, CHI ’24, Association for Computing Machinery, New York, NY, USA, 2024. URL: https: //doi.org/10.1145/3613904.3642867. doi:10.1145/3613904.3642867. [23] H.-S. Chang, H.-J. Hsu, K.-T. Chen, Modeling exercise relationships in e-learning: A unified approach, in: Educational Data Mining, 2015. URL: https://www.educationaldatamining.org/ EDM2015/proceedings/short532-535.pdf. [24] H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y. Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosale, et al., Llama 2: Open foundation and fine-tuned chat models, 2023.
URL: https://arxiv.org/abs/2307.09288. arXiv:2307.09288. [25] A. Grattafiori, A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Vaughan, et al., The llama 3 herd of models, 2024. URL: https://arxiv.org/abs/2407. 21783. arXiv:2407.21783.