Feedback plays an important role in education. However, it is sometimes limited by the amount of time the instructor can dedicate to reviewing each student's work. To address this limitation, this article presents ReflectIA, an AI-based evaluation system that combines criteria-based visual analysis with dialogue with a student-facing dialogue system designed to structure reflective interaction in graphic design. The system analyzes the images provided by the student to obtain an overall understanding of the creative design project's content. The student can interact with the system via a conversation that provides feedback through question-based tutoring aligned with the principles of reflective practice. ReflectIA was tested with a group of 58 third-year undergraduate engineering students working on two design assignments: an Editorial Magazine and a Digital Campaign. The dialogues between students and the system were analyzed to characterize the interaction patterns, linguistic and semantic properties, and pedagogical functions that emerged during the feedback. Results show that the system generated question-oriented feedback and interaction patterns consistent with a reflective critique process. This study did not assess improvements in student learning or design quality. These findings suggest that AI-mediated feedback can complement instructor-led critique by offering students structured and individualized opportunities to examine and discuss their design decisions, especially when it is difficult to provide ongoing feedback. More broadly, the results show that multimodal AI can generate observations about design work and organize them into a question-based critique, creating opportunities to develop reflection-oriented classroom activities.
To accommodate diverse learner needs, personalized learning requires educational resources that adapt in both content complexity and mode of presentation. Existing studies indicate multi-modal content facilitates easier understanding of complex concepts that are difficult to communicate through text alone. However, creating pedagogically appropriate resources demands substantial time and effort, limiting their scalability. This work presents BioVisGen, an agentic multi-modal framework for educational diagram generation from textual descriptions, integrating specialized agents for prompt structuring, image generation, output evaluation, and refinement. To enable domain-specific adaptation, the framework is built upon OmniGen2 and fine-tuned on BioKosh, an in-house dataset comprising 17,000 biological images paired with three levels of textual descriptions representing varying pedagogical complexity. The framework further incorporates assessment and refinement mechanisms to improve semantic consistency and visual quality. In addition, we propose an education-specific qualitative evaluation framework for assessing the generated images through human and AI evaluations. Experimental results demonstrate that BioVisGen outperforms domain-adapted GAN- and diffusion-based baselines and achieves visual quality comparable to contemporary vision-language models while being trained on a relatively small dataset and requiring lower computational resources.
Multi-User Virtual Environments (MUVEs) have gained increasing attention in collaborative learning research due to their interactive affordances. However, most current MUVEs remain desktop-based, limiting their ability to simulate authentic learning contexts. Immersive virtual reality (IVR), delivered through head-mounted displays that engage learners' peripheral vision, can enhance their sense of presence and immersion within virtual environments. To explore this potential, the present study proposes a multidimensional design framework that encompasses social, motivational, cognitive, and experiential dimensions and uses it to develop an Immersive Multi-User Virtual Environment (MUIVE). The effectiveness of MUIVE was evaluated using measures of learning performance, social presence and collaboration quality (social), learning motivation (motivational), cognitive load (cognitive), and IVR experience, trust, ease of use, and satisfaction (experiential). Results showed a higher overall performance gain for MUIVE in the gain-score analysis, although the corresponding baseline-adjusted analysis did not yield a statistically significant condition effect. MUIVE significantly enhanced social presence, learning motivation, trust, ease of use, and satisfaction, without increasing cognitive load. Based on these findings, we propose five core design principles for building effective MUIVEs: (a) face-to-face authenticity, (b) engagement enhancement, (c) extraneous cognitive load minimization, (d) avatar embodiment, and (e) system experience. Furthermore, peer familiarity moderated the effect of platform on transfer performance, with MUIVE showing a stronger advantage among unfamiliar dyads. These findings offer design implications for the development of immersive collaborative learning systems and serve as a foundation for future research on optimizing multi-user interaction and learning support in IVR-based environments.
Whether physical social robots benefit student learning within computer-based learning environments is unclear. We used a pretest and posttest design with random assignment to compare a classic computer-based learning environment featuring on-screen teacher and peer agents with an otherwise identical version in which both agents were also present as physical robots. Outcomes were learning gains, topic-related intrinsic value, and students' ratings of each agent's sociability. We expected embodiment to increase learning, value, and perceived sociability. We also contrasted two potential pathways by comparing a teacher agent enriched with human-like communicative cues and a small toy-like peer agent designed to leverage cuteness and low status. Contrary to predictions, embodiment did not increase learning gains or topic-related intrinsic value. However, students rated the peer robot as more sociable in the robotic condition, whereas ratings of the teacher robot did not differ by condition. Across conditions, higher perceived sociability for either agent was associated with higher posttest topic-related value after controlling for pretest scores, while no parallel effects on knowledge were observed. Planned moderation by condition and mediation through perceived sociability were not supported. The findings indicate that the physical presence of a small toy-like peer robot can raise perceived sociability, yet embodiment alone is unlikely to shift learning or value. In the present study, this increase in perceived sociability was not accompanied by corresponding gains in learning or topic-related value, highlighting the need to establish the educational added value of physical embodiment over screen-based alternatives.
Personalized lesson plan generation remains a critical challenge in adaptive learning, where instructional content must be tailored to the evolving ability profiles of individual learners. Existing LLM-based frameworks, such as EduPlanner [1], adopt a dual-agent evaluation-optimization loop but lack dynamic student modeling, adaptive question selection, and self-refinement mechanisms, limiting their effectiveness across diverse educational domains. In this paper, we propose EduPlanner+, the LLM-based multi-agent framework to jointly integrate dynamic student ability evolution, self-reflective lesson refinement, and weak-ability-targeted question selection for personalized lesson plan generation in multi-domain adaptive learning. EduPlanner+ orchestrates five components under the Optimizer Agent Expert Optimization pipeline: (1) the Evaluator Agent Expert Evaluation for multi-expert lesson scoring, (2) the Analyst Agent Expert Analysis for error-prone point augmentation, (3) the Adaptive Question Selector for weak-ability-targeted evaluation, (4) the Dynamic Skill Tree for evolving student persona modeling, and (5) the Reflection Agent for self-critique-based lesson refinement. Experiments on four benchmark datasets: Algebra, GSM8K, Database System, and Operating Systems, demonstrate that EduPlanner+ consistently outperforms state-of-the-art across all domains, achieving superior evaluator agent expert evaluation scores on every evaluated dataset. Ablation studies confirm that the Dynamic Skill Tree, Reflection Agent, and Adaptive Question Selector are the most impactful components for personalized lesson quality improvement.
