Personalized learning recommendations utilizing learning behavior patterns have emerged as a crucial research area. Backtracking behavior, defined as students revisiting pages that they have already studied, has been shown to improve learning outcomes. However, it is difficult for students to locate the content they wish to revisit within a large number of resources. In this study, we designed a page-jumping recommendation plugin to help students find the content they require, thereby enhancing their backtracking behavior. By analyzing educational content and student reading logs, this plugin can predict and suggest pages that students are likely to want to revisit. An experiment was conducted in a university course to verify the plugin's effectiveness with respect to learning outcomes. Students who used the page-jumping recommendation plugin achieved better academic performance and demonstrated high acceptance of the system.
Recent advances in image-generative artificial intelligence (AI) have provided designers with powerful tools for quickly exploring and iterating on their creative concepts. However, the effectiveness of different sources of inspiration in supporting design development remains underexplored, particularly within design education. This study examines the impact of AI-generated images versus web-sourced images on designers' creative outcomes. For this purpose, we have gathered a dataset of design projects created by 50 students, who used AI-generated and web-sourced images as inspiration. To identify sources of inspiration in each project, we propose computer vision metrics that measure contrast and color similarity, as well as deep learning embedding similarity. The similarities computed between final designs and inspiration images are compared with student surveys expressing their perception of sources of inspiration and with expert ratings assigned to each project. The comparison shows that the students' stated inspiration is positively correlated with the computed inspiration-design similarity and that students exhibit a preference for AI-generated inspiration, even if they are first-time users of generative tools. Moreover, our study reveals that both web and AI images are inspirational, but they affect different aspects of design. AI inspirations correlate strongly with the overall composition and mood, and web-sourced inspirations play an important role in defining color, materials, and grounding the design in reality.
In recent years, the rise of detailed educational data has fueled the growth of learning analytics (LA), particularly in blended learning. While LA provides the methodology for analyzing course features and predicting performance, a robust theoretical lens is needed to interpret student behaviors meaningfully. To better understand students’ digital learning behaviors, it is essential to integrate educational theory with data-driven methods by combining behavioral data with self-reported information. Consequently, this study addresses this by investigating the relationship between interaction behaviors, learning styles, and learning outcomes grounded in Self-Regulated Learning (SRL) theory. A total of 233 first-year university students participated in a course integrating a digital textbook system, namely Digital Textbook for Improving Teaching and Learning (DITeL). Behavioral interaction data and self-reported learning styles were analyzed using correlation analysis and hierarchical clustering. Three distinct behavioral patterns were identified, with the deep engagement group outperforming others in GPA. Reflective and sensing learners demonstrated more active interaction with the digital system. The findings provide valuable insights into the dynamics of LA and student behavior analysis in blended learning models.
Research has highlighted the importance of self-regulated learning (SRL) for personal development and lifelong learning. Therefore, it is essential to foster SRL skills in learners from a young age to prepare them to adapt to and succeed in a rapidly changing world. Generative AI (GenAI) tools offer significant potential for helping learners develop these skills through immediate, individualised feedback. This study evaluated the use of GenAI learning tools to enhance students’ SRL skills. Eighty-four senior secondary students completed a 30-hour GenAI application course and were required to reflect on their SRL skills. A mixed-methods approach was employed to evaluate the course. Pre- and post-survey comparisons indicated that the students’ SRL skills improved significantly after attending the course. A thematic analysis of their self-reflective writings revealed that the students strongly believed the course had inspired them to apply GenAI tools for acquiring subject knowledge and managing their learning. The effectiveness of this course can guide and inspire future empirical research and pedagogical practices focused on integrating GenAI tools into the educational process to develop students’ SRL.
Generative AI produces convincing work quickly, but this convenience breaks the link between a finished product and its author’s capability. The central problem is no longer whether AI should be used, but which parts of a task should be performed, learned, or delegated. While current frameworks in automation, self-regulated learning, cognitive offloading, and human-AI interaction illuminate facets of this problem, none offer a systematic method for task allocation that balances immediate workflow needs with the preservation of long-term human capability. A framework is needed to optimize work distribution while safeguarding long-term human capability and value. This paper introduces the Know-Learn-Delegate (KLD) framework, a capability-based solution that treats the relation between a person, a task, and that person’s developmental intention as the unit of allocation. Users decompose a task and classify each component according to whether they can perform it unaided and, if not, whether they intend to learn it. AI then serves as proxy executor, tutor, or autonomous agent. KLD’s originality lies in connecting immediate efficiency with capability formation, human agency, capability maintenance, and assessment. To prevent inflated claims of competence, a three-part “Know” test comprising Perform, Specify, and Detect (PSD) is introduced. Delegation is further constrained by state-specific application of Verifiability, Consequence, and Reversibility (VCR) principles. This framework also mandates ongoing practice for essential capabilities and establishes stronger restrictions during formative learning stages. Design principles translate KLD into multi-agent workflows in which planning, generation, evaluation, and human acceptance remain essential. The paper proposes assessing human-AI contribution through observable agency across the production process rather than percentages inferred from the final artifact. KLD thus offers educators, learners, and knowledge workers a defensible way to gain AI’s efficiency without quietly surrendering human value in terms of agency, judgment, or future capability.
