We’re truly humbled and grateful to share our latest collaborative research, published in Education and Information Technologies (2026). Titled “Exploring the relationship between computational thinking behaviors and socially shared metacognitive regulation in collaborative coding across task complexity”, this work represents a meaningful partnership between Hanyang University in South Korea and Texas A&M University in the USA. As part of this cross-continental effort, we delved into how non-computer science (non-CS) major undergraduates navigate collaborative block-based coding, uncovering patterns that could shape the future of AI-enhanced learning environments.
Working alongside lead author Dr. Yoonhee Shin, Seohyun Choi, and Sujung Lee from Hanyang University, Dr. Song’s contribution at Einbrain Lab focused on processing screencaptured video data from learners’ coding sessions. Using machine learning techniques, specifically a convolutional neural network (CNN) for supervised classification of block types and unsupervised clustering to identify behavioral profiles, we extracted and analyzed patterns in computational thinking (CT) behaviors. This approach allowed us to reveal how high-performing groups achieved more efficient and coherent regulatory cycles, especially in complex tasks. It’s been an inspiring journey, and I’m thankful for the opportunity to collaborate with such talented colleagues.
Why This Research Matters: Background and Motivation
As computational thinking (CT) emerges as a core skill across disciplines, integrating it into education for non-CS majors has become essential. CT involves problem-solving using principles like abstraction, algorithmic thinking, and debugging, often taught through programming. However, novices frequently struggle with cognitive overload, fragmented reasoning, and inefficient trial-and-error approaches, particularly in collaborative settings.
Block-based coding tools like MIT App Inventor lower these barriers by reducing syntactic complexity, allowing learners to focus on conceptual structures. Yet, prior studies have emphasized coding outcomes over the underlying processes, especially how Socially Shared Metacognitive Regulation (SSMR), the group’s collective planning, monitoring, and adaptation, interacts with CT behaviors across varying task difficulties.
Our study addresses these gaps through a mixed-methods lens, combining behavioral video analysis with discourse examination. We posed two key research questions:
- RQ1: How do CT behavioral patterns in collaborative coding vary by task complexity and problem-solving performance?
- RQ2: How do SSMR patterns unfold in relation to these CT behaviors, and how do they differ by performance and complexity?
This research aligns perfectly with Einbrain Lab’s mission to advance XR (VR/AR/MR) solutions integrated with AI for human-AI collaboration in technical training. By understanding these patterns, we can design immersive XR environments that provide adaptive scaffolding and real-time feedback, fostering innovation in computational education.
Methodology: A Closer Look at Our Approach
We conducted the study with 120 undergraduate freshmen from a social sciences major at a large South Korean university, enrolled in an introductory CS course. Participants were randomly grouped into 30 teams of four, emphasizing collaborative problem-solving.
The course spanned seven weeks in an online format:
- Weeks 1-4: Asynchronous lectures on CT fundamentals via the university’s LMS.
- Week 4: Pre-test to assess prior knowledge.
- Weeks 5-7: Synchronous collaborative coding sessions using Zoom and App Inventor, with rotating “driver” roles for screen sharing.
Tasks increased in complexity:
- Foundational (Week 5): Simple math calculator (sequences and events).
- Basic (Week 6): BMI calculator (conditionals and local variables).
- Advanced (Week 7): Course registration system (global variables, lists, loops).
We collected 60 video recordings (from Weeks 6-7) for analysis, focusing on active coding segments totaling over 114,000 seconds.
Behavioral Analysis with Machine Learning
Dr. Song’s role centered here: We classified blocks as computational concept (CC) blocks (e.g., control, logic, variables) or non-CC blocks (e.g., math, text). A CNN model, trained on manually annotated samples, achieved high accuracy (F1-scores: 0.86 for CC, 0.93 for non-CC). This supervised classification automated block identification across videos.
Unsupervised K-means clustering then grouped behaviors based on total blocks, CC/non-CC counts, and transitions, revealing three clusters (A, B, C) differentiated by efficiency and selectivity.
Discourse Analysis
Group discussions were coded using an SSMR framework (orienting goals, making plans, enacting strategies, monitoring/controlling, evaluating/reflecting, adapting metacognition). Lag sequential analysis identified significant transitions, stratified by performance (top/bottom 15%) within clusters.
Key Findings: Patterns in CT and SSMR
CT Behavioral Patterns (RQ1)
Clustering revealed distinct profiles:
- Cluster A (Extensive-Execution): High total blocks (mostly non-CC), moderate performance.
- Cluster B (Outlier): Very high blocks, limited to basic tasks.
- Cluster C (Efficiency-Oriented): Fewer total/non-CC blocks, higher CC selectivity, superior performance, especially in advanced tasks.
| Cluster | Total Blocks (Mean, Basic Task) | Non-CC Blocks | CC Blocks | PS Performance |
|---|---|---|---|---|
| A | 1996.8 | 1811.1 | 185.7 | 5.6 |
| B | 4135.5 | 3478.8 | 656.8 | 6.0 |
| C | 1447.0 | 1015.6 | 431.4 | 6.1 |
High performers (Cluster C) demonstrated concise, strategic block usage, avoiding unnecessary tinkering.
SSMR Patterns (RQ2)
Discourse revealed performance-linked regulatory sequences:
- Successful Groups (Cluster C): Coherent cycles like ER→AP (evaluation to adaptation), sustaining efficiency in complex tasks.
- Unsuccessful Groups (Cluster A): Reactive patterns, e.g., ES→OG (enactment before goal-setting), leading to fragmented regulation.
In advanced tasks, high performers adapted proactively, while low performers looped in superficial planning-evaluation without action.
Implications for Education and Einbrain Lab’s Work
These findings highlight that success in collaborative coding stems not from volume but from selective CT application and robust SSMR. Low performers rely on trial-and-error, while high performers exhibit concise, adaptive cycles.
For Einbrain Lab, this opens doors to AI-driven XR tools:
- Adaptive Scaffolding: Real-time monitoring of block patterns and discourse could trigger prompts, e.g., “Reflect on your evaluation: how can you adapt?” in VR environments.
- Human-AI Collaboration: XR simulations could integrate AI agents that co-regulate group processes, providing feedback on regulatory gaps.
- Technical Training: In immersive MR settings, learners could practice complex tasks with AI-guided regulation, enhancing innovation in fields like engineering.
Future work might validate these in XR contexts or incorporate multimodal data (e.g., physiological signals) for deeper insights.
Final Thoughts
This collaboration has been a profound learning experience, reinforcing how diverse perspectives drive meaningful advancements. Grateful to Dr., Yoonhee Shin for her leadership, let’s continue pushing boundaries!
Article link: https://link.springer.com/article/10.1007/s10639-026-13917-1
#ComputationalThinking #AIinEducation #CollaborativeLearning #XR #HumanAICollaboration #EinbrainLab #InternationalResearch