This paper explores the role of generative AI (GenAI) in providing adaptive summative feedback within an embodied learning environment for children's mathematics education. Using a body-scale digital number line, children engaged in learning integer operations through physical interaction. The study employed a betweengroup design: one group received feedback from a human instructor, while the other received AI-generated feedback. A mixed-method approach combined multimodal data (system logs, motion sensors) with qualitative observations of student interactions. The results showed no significant diferences in task performance but revealed key diferences in engagement: the teacher feedback encouraged multimodal, reflective responses involving gestures and body movements, while the AI feedback promoted streamlined, task-specific strategies with reduced cognitive load. These findings demonstrate the complementary strengths of human and AI feedback, underscoring the potential of hybrid intelligence systems to enhance adaptive learning environments.
Embodied learning, rooted in embodied cognition theory, links cognitive processes with physical
interactions in the environment [
RQ: How do teacher and GenAI feedback difer in shaping cognitive engagement, task eficiency, and multimodal interaction in embodied learning environments, and what is the potential for hybrid integration (with a focus on expanding teachers’ skills and not replacing them)?
By examining the benefits and limitations of GenAI feedback, this study contributes to understanding
its role in education, highlighting implications for educators, researchers, and developers aiming to
design hybrid intelligence systems that enhance learning experiences [
Embodied cognition theory emphasizes the link between bodily experiences and cognitive processes,
highlighting the role of physical actions such as gestures and movements in facilitating learning
[13, 14, 15]. In mathematics education, embodied learning approaches help bridge abstract concepts
with practical understanding, significantly improving comprehension and retention [
This study utilized a portable platform called MOVES designed to enable MSEs, addressing limitations
inherent in traditional learning environments. The educational framework involved students engaging
with a body-scale NL to solve integer arithmetic problems. This NL served as an efective teaching
tool, presenting integers in a spatially organized and ordinal layout, where negative integers mirrored
their positive counterparts [
In Figure 1 an example of the NL activity. To solve the equation “-1 – 2,” the student begins by standing on the -1 hash mark on the NL. As they step onto -1, the number beneath their feet turns blue, while the -1 displayed on the wall in front changes to green, accompanied by a sound. Next, the student turns to the left to face the subtraction direction. Upon turning, the subtraction sign on the wall also turns green, and another sound plays. Finally, the student takes two steps forward and raises their hands above their head to indicate that they have reached the solution. If correct, the entire problem, along with the solution on the wall, turns green, and a congratulatory sound plays. If incorrect, the solution remains unchanged, and no sound plays, as negative feedback is avoided to prevent demotivating or discouraging students [29].
This study integrated the GPT-4 language model into a Node.js web server to enhance the interactive educational systems reliant on body movement and dynamic user input. The system uses real-time sensor data from students interacting with the number line, sending responses to the web server, which communicates with GPT-4 via OpenAI’s API to generate personalized feedback. The Node.js server ensures eficient asynchronous communication, incorporates caching to reduce redundant API calls, and includes error-handling protocols for reliable operation. To ensure the efectiveness of GenAI feedback, the prompts were co-designed with the teacher, aligning AI responses with specific learning objectives and instructional strategies. This iterative process created a seamless blend of AI-driven feedback and human pedagogical guidance, delivering contextually relevant, constructive suggestions to support student engagement and learning. Below, we present an example of feedback generated by GPT-4 from system logs.
GenAI FEEDBACK EXAMPLES FROM LOGS: EXAMPLE: NLS:operationSelected|"1|+|-4" NLS:LLM-UserMessage|"\nO1: +1\nOperator: + \nO2: -4 \nResult: -3\n The student walked on the correct number: 1\nThe student correctly rotated his body to the right\nThe student correctly walked backward\n The student walked on the correct number: -3\nProvide feedback." NLS:LLM-Response|"You correctly walked to 1, rotated right, and walked back 4 steps to -3. Well done following the operation rules with negative numbers on the number line!"
In collaboration with a secondary school in Trondheim, the study lasted 2 weeks in May 2024. The teacher provided us with a list of participants for each day, and one of the researchers randomly called the participating students from the class list one by one; on average, each student’s session lasted 40-50 minutes. The study aimed to assess the impact of adaptive summative feedback within the MOVES-NL educational tool. Participants were divided into two groups, a control Group (N=16) that received summative feedback from a teacher and a experimental Group (N=14)that received summative feedback through GenAI. The study room was set up in a dedicated classroom inside the school to avoid external distractions. Each study session consisted of the phases: 1. Facilitator’s introduction covering our identity, planned activities, and the data collection process (including camera recording and eye-tracking setup). 2. Introduction to the Walking NL: the facilitator showed students how to move to solve problems. 3. Walking NL: the student walks the NL in order to solve arithmetic problems while an avatar mirrors movements on the screen. 4. Summative feedback: the student gets feedback on their task based on the group setting.
During the sessions, children worked on various dificulty levels of math problems, with specific success criteria outlined for each level.
Our sample consisted of 34 students (18 females) between 11-13 years old, selected based on the curriculum timing when students begin learning about negative numbers and their applications in mathematical concepts. Prior to their participation, written informed consent was obtained from their legal guardians. All the ethical procedures were approved by the national human research ethics organization. Students’ participation and data collection were conducted after approval from the Norwegian Agency for Shared Services in Education and Research (Sikt) and with Institutional Review Board approval (protocol ID 2022-10-15703), following all the regulations and recommendations for research with students.
We recorded students’ interactions with MOVES-NL and employed sensing devices which allowed us to capture students’ experience via multimodal data. The decision to use these data collection techniques was also influenced by the fact that they account for (to some extent) students’ embodied learning and their importance in multisensory systems (e.g., students externalize their actions with the use of their body/skeleton). The sensing devices and their respective multimodal data allowed us to closely monitor and understand how students experienced the received support, leveraging the key afordances of multimodal data (e.g., temporality and direct access to indicators of students’ cognitive and afective processes [30]). Students’ activity sessions were recorded using two mobile cameras and one additional sensor device: gaze data from eye-tracking glasses. We collected children’s gaze data using Tobii eye-tracking glasses at a 50 Hz sampling rate and one-point calibration. Due to errors in data collection, we had to discard four students.
The results reveal distinct patterns in how students interacted with feedback in the teacher and GenAI conditions, informed by both quantitative metrics (system logs, eye tracking, and motion sensors) and qualitative observations (video analysis). While task performance did not difer significantly between the two groups ((31.63) = 0.73, > 0.05 ), notable diferences emerged in cognitive engagement and interaction styles. Students in the teacher feedback condition exhibited higher cognitive load ((26.12) = 2.89, < 0.01 ) and employed global processing strategies, as evidenced by their higher Information Processing Index (IPI) scores ((27.56) = 5.16, < 0.001 ). These students frequently re-enacted embodied solutions to problems, using gestures and full-body movements to justify their answers, reflecting the multimodal nature of human interaction. In contrast, students in the GenAI feedback condition experienced lower cognitive load, focusing on task-specific verification strategies that prioritized correct responses rather than reasoning behind the incorrect ones. Their interactions were less physically expressive, with fewer gestures observed during feedback responses, as students concentrated on detecting whether the system’s evaluation was accurate. Time spent on Areas of Interest (AOIs) further highlights these diferences: students in the teacher condition focused on feedback and task text ((22.64) = 5.06, < 0.001 ), while those in the GenAI condition engaged more with the correct option and the number line ((25.03) = 2.68, < 0.05 ). These findings suggest that the teacher feedback encouraged a more reflective, multimodal engagement with problems, while the GenAI feedback streamlined the learning process, fostering eficiency but with reduced physical expressiveness. Together, these insights underline the complementary strengths of human and AI feedback in educational contexts, ofering opportunities to balance critical reflection with task-focused eficiency.
The preliminary results highlight distinct benefits and limitations of teacher and GenAI feedback within an embodied learning context. Teacher feedback encouraged multimodal interaction and critical reflection, as evidenced by students’ frequent use of gestures and body movements to justify their answers. This aligns with higher cognitive load and global processing strategies, suggesting that teacher feedback prompts deeper engagement and fosters metacognitive skills. However, this also comes at the cost of increased mental efort, which may impact eficiency during problem-solving tasks. Conversely, GenAI feedback reduced cognitive load and supported eficient, task-focused learning. Students in the GenAI condition concentrated on verifying the accuracy of feedback and interacting with the number line, indicating streamlined engagement with less emphasis on multimodal expression. While this eficiency can enhance learning for routine tasks, it may limit opportunities for critical reflection and the development of higher-order thinking skills.These findings suggest that hybrid intelligence systems combining teacher and AI feedback could leverage the strengths of both approaches. Teachers can provide reflective, multimodal engagement to deepen understanding, while GenAI can ofer consistent, task-specific feedback and reduce cognitive demands, and enhance eficiency. Such integration could create adaptive learning environments that cater to diverse student needs and learning objectives. Future research should explore ways to balance these complementary strengths, focusing on designing hybrid systems that integrate AI feedback seamlessly with human instruction. This includes addressing the observed limitations of GenAI, such as reduced physical expressiveness and student skepticism. It´s integration in educational settings necessitates a consideration of critical ethical dimensions. Data privacy is a paramount concern, as student interactions and personal data must be protected under stringent regulations. Additionally, potential biases in AI-generated feedback could inadvertently reinforce inequalities, emphasizing the need for diverse and inclusive training datasets. Lastly, the emotional impact of AI interactions on students must be carefully managed to avoid demotivation, underscoring the importance of designing systems that complement AI eficiency with human-centered support. Addressing these aspects is essential for the responsible deployment of GenAI in education. Longitudinal studies could investigate the long-term efects of hybrid feedback on learning outcomes and student engagement and learning trajectories. Coupled with qualitative insights from teachers and learners we could explore how sustained exposure to AI and teacher feedback influences critical thinking, problem-solving skills, and overall academic growth with a comprehensive understanding of the enduring impacts of hybrid intelligence systems in education. The hybrid model promoted by this research has significant implications for teacher training and professional development. To efectively collaborate with AI systems, teachers need skills in interpreting AI-generated feedback, integrating it with their instructional strategies, and fostering student engagement through complementary human insights. Professional development programs should incorporate training on these aspects, including hands-on experiences with AI-driven educational tools, workshops on aligning AI feedback with pedagogical goals, and strategies for adapting to diverse student needs in hybrid learning environments. By equipping teachers with these competencies, educational institutions can maximize the potential of hybrid intelligence systems to enhance learning outcomes. Furthermore, the scalability of these systems beyond controlled environments is a critical consideration. Implementing such AI-driven educational tools in diverse classroom settings requires addressing challenges such as infrastructure availability, teacher training, and adaptability to resource-limited or underserved contexts. In conclusion, the study underscores the potential of hybrid feedback systems in embodied learning environments. By combining the teacher’s contextual understanding and multimodal engagement with the AI’s eficiency and adaptability, hybrid approaches can enhance the educational experience, promoting both critical reflection and task-focused learning.
We would like to thank all the children and teachers who participated in this study. Without their enthusiasm and dedication, this work would not have been possible. Moreover this study was possible thanks to the Peder Sather Grant programme. [13] P. Dourish, Where the action is, MIT press Cambridge, 2001. [14] J. Piaget, Genetic epistemology, Columbia University Press, 1970. [15] D. Abrahamson, A. Bakker, Making sense of movement in embodied design for mathematics learning, Cognitive research: principles and implications 1 (2016) 1–13. [16] M. M. Riconscente, Results from a controlled study of the ipad fractions game motion math, Games and Culture 8 (2013) 186–214. [17] R. Barrocas, J. Bahnmueller, S. Roesch, M. Lachmair, K. Moeller, Design and empirical evaluation of a multitouch interaction game-like app for fostering early embodied math learning, International Journal of Human-Computer Studies 175 (2023) 103030. [18] G. Cosentino, M. Giannakos, Multisensory interaction and analytics to enhance smart learning environments: A systematic literature review, IEEE Transactions on Learning Technologies (2023). [19] N. Sinclair, E. Heyd-Metzuyanim, Learning number with touchcounts: The role of emotions and the body in mathematical communication, Technology, Knowledge and Learning 19 (2014) 81–99. [20] M. Chen, Small leans into big steps: A mixed-reality environment to support embodied, ensembled mathematics learning, in: Interaction Design and Children, 2022, pp. 564–568. [21] S. A. Papert, Mindstorms: Children, computers, and powerful ideas, Basic books, 2020. [22] W. Yang, X. Hu, I. H. Yeter, J. Su, Y. Yang, J. C.-K. Lee, Artificial intelligence education for young children: A case study of technology-enhanced embodied learning, Journal of Computer Assisted Learning 40 (2024) 465–477. [23] S. Druga, S. T. Vu, E. Likhith, T. Qiu, Inclusive ai literacy for kids around the world, in: Proceedings of FabLearn 2019, 2019, pp. 104–111. [24] Y. Huang, Y. Cheng, L. Chen, J. Y. Hsu, Human-ai co-learning for data-driven AI, CoRR abs/1910.12544 (2019). http://arxiv.org/abs/1910.12544. [25] S. Järvelä, J. Malmberg, H. Järvelä, Generation z and beyond: Co-evolution of human capabilities and intelligent technologies in the twenty-first century (2022) 1–13. doi: 10.1007/ 978- 981- 19- 6887- 7_115. [26] S. Järvelä, A. Nguyen, A. Hadwin, Human and artificial intelligence collaboration for socially shared regulation in learning, British Journal of Educational Technology (2023). doi:10.1111/bjet.13325. [27] G. Cosentino, J. Anton, K. Sharma, M. Gelsomini, M. Giannakos, D. Abrahamson, Hybrid teaching intelligence: Lessons learned from an embodied mathematics learning experience, British Journal of Educational Technology (2024). [28] M. Gelsomini, Sensei - experience in all senses. multisensory and multimodal research framework, 2023. https://sensei.space/. [29] A. C. Van Duijvenvoorde, K. Zanolie, S. A. Rombouts, M. E. Raijmakers, E. A. Crone, Evaluating the negative or valuing the positive? neural mechanisms supporting feedback-based learning across development, Journal of Neuroscience 28 (2008) 9495–9503. doi:10.1523/JNEUROSCI.1485- 08. 2008. [30] M. Cukurova, M. Giannakos, R. Martinez-Maldonado, The promise and challenges of multimodal learning analytics, Wiley Online Library (2020). doi:10.1111/bjet.13015.