From Passive Consumption to Active Interaction: Exploring Interactive LLM Scaffolding to Support Learning Engagement

This paper presents a small-scale laboratory study demonstrating that embedding lightweight interactive components into Large Language Model-generated scaffolding can shift learners from passive consumption to active engagement, thereby improving perceived attention and short-term learning outcomes.

The Gist

Imagine you are trying to learn a complex magic trick. You have a wizard (the AI) who can explain exactly how it's done.

The Old Way (Passive Consumption):
In the current version of AI tutoring, the wizard just hands you a long, written scroll with the instructions. You sit there, read it, and nod. It's like watching a cooking show where the chef does everything, and you just watch. You might understand the words, but your brain is on "autopilot." You aren't really doing the cooking; you're just watching the recipe.

The New Idea (Active Interaction):
This paper asks: What if, instead of just reading the scroll, you had to physically uncover the instructions yourself?

The researchers built a prototype where the AI's helpful hints and explanations are initially hidden under a digital "scratch-off" layer (like a lottery ticket or a scratch-and-win game). To see the AI's advice, you have to use your mouse to "scratch" away the gray cover.

The Experiment

The team tested this with 8 computer science PhD students learning difficult math proofs. They split the learning into two modes:

1. The "Plain Text" Mode: The AI just types out the answer (like the old scroll).
2. The "Scratch-Off" Mode: The AI's answer is covered up. The student has to actively scratch it off to read the hint.

What They Found

Even though the study was small, the results were surprising and encouraging:

• It didn't feel like more work: You might think scratching a screen would be annoying or tiring. But the students actually felt less frustrated and found the interface easier to use. It's like how fidgeting with a stress ball can actually help you focus better than just sitting still.
• Better Focus: Because they had to "do" something to get the answer, their brains were more alert. They weren't just skimming; they were engaging.
• Slightly Better Grades: The students who used the scratch-off method scored slightly higher on their quizzes. It seems that the tiny physical effort of "uncovering" the knowledge helped lock it into their memory.
• Fewer Questions: Interestingly, students in the interactive mode asked the AI fewer follow-up questions. It seems that by engaging with the hints they did get, they understood the material faster and didn't need to ask for more help.

The Bigger Picture: A Toolbox, Not a Hammer

The researchers didn't stop at just "scratching." In follow-up interviews, they brainstormed other ways to make AI learning more active. They realized that different types of help need different "interactive tools":

• For Explanations: Instead of scratching, maybe you hover your mouse over a hard word to see a quick definition pop up (like a tooltip).
• For Step-by-Step Guides: Instead of reading a list, maybe you have to drag and drop the steps into the correct order, like solving a puzzle.
• For Modeling: Maybe you can unfold a complex diagram piece by piece, so you aren't overwhelmed by seeing the whole picture at once.

The Takeaway

The main lesson is simple: Learning is like exercise. If you just watch someone else lift weights, you don't get strong. You have to lift the weights yourself.

Currently, AI tutors often just "talk" at us. This paper suggests that by adding tiny, lightweight interactions (like scratching, clicking, or dragging), we can turn AI from a passive narrator into an active gym partner. It forces the learner to stop and think, turning "reading" into "doing," which leads to better learning.

Technical Summary

Here is a detailed technical summary of the paper "From Passive Consumption to Active Interaction: Exploring Interactive LLM Scaffolding to Support Learning Engagement."

1. Problem Statement

The paper addresses a critical limitation in current Large Language Model (LLM)-based learning systems: passive consumption.

• Context: LLMs are increasingly used as personal tutors to provide scaffolding (hints, explanations, step-by-step guidance).
• The Issue: Current implementations rely on standard chatbot interfaces where learners primarily read static text outputs. This contrasts with established learning science theories (e.g., the ICAP framework), which posit that learning outcomes depend heavily on the mode of engagement. Passive reading yields lower cognitive engagement compared to active, constructive, or interactive engagement.
• The Gap: While LLMs can generate high-quality scaffolding, the delivery mechanism (static text) fails to actively involve the learner with the content itself. The challenge is not what the LLM generates, but how that content is presented to force active engagement.

2. Methodology

The authors conducted a within-subjects laboratory study (N=8) to empirically test whether embedding lightweight interactive components into LLM responses improves engagement and learning.

A. System Prototype (Design Probe)

The researchers built an LLM-based tutoring system with two distinct conditions:

1. Non-Interactive Condition (Control): The LLM provides standard plain-text responses containing scaffolding (definitions, hints, guidance).
2. Interactive Condition (Experimental): The LLM generates the same content, but specific scaffolding segments are initially masked. Learners must perform a scratch-off interaction (mouse-based) to reveal the hidden text. This forces the learner to physically interact with the content before accessing it.

• Technical Implementation: The system uses the GPT-5.2 model. Prompts are tailored to include established scaffolding practices. A client-side script detects LLM-marked segments and applies the masking/reveal logic.
• Task: Participants studied two college-level number theory proofs (adapted from prior research) and answered comprehension quizzes.

B. Experimental Design

• Participants: 8 Computer Science PhD students with undergraduate math backgrounds.
• Procedure:
  • Counterbalancing: Participants learned one proof under the interactive condition and the other under the non-interactive condition (order and proof assignment were counterbalanced).
  • Phases: Exposure (skimming) $\rightarrow$ Learning (15 mins max per proof with LLM) $\rightarrow$ Quiz (10 mins, 10 MCQs) $\rightarrow$ Post-study survey $\rightarrow$ Think-aloud interview.
• Metrics:
  • Learning Outcomes: Quiz scores (comprehension of proof structure and logic).
  • Engagement/Workload: NASA-TLX (workload), System Usability Scale (SUS), and Likert-scale surveys on perceived engagement and focus.
  • Behavioral: Number of questions asked to the LLM.
  • Qualitative: Think-aloud interviews to gather design ideas for future interaction.

3. Key Contributions

1. Conceptual Shift: Proposes a shift from viewing LLM scaffolding as static text delivery to interactive content delivery. It argues that interaction serves as a mechanism to direct attention and enforce active processing.
2. Design Probe: Introduces a specific, lightweight interaction pattern (scratch-off masking) that successfully transforms passive reading into an active retrieval/reveal process without requiring complex UI changes.
3. Pedagogical Mapping: Through the think-aloud study, the authors map specific interaction types to distinct scaffolding means (Hinting, Instructing, Explaining, Modeling, Questioning, Feedback), providing a framework for future design (e.g., using drag-and-drop for sequencing, hover-annotations for explaining).

4. Results

Despite the small sample size (N=8), the study yielded promising initial evidence:

• Learning Outcomes: The interactive condition showed a positive trend in quiz performance ($M=80.0\%$) compared to the non-interactive condition ($M=77.5\%$), though the difference was not statistically significant due to the sample size.
• Engagement & Focus: Participants reported significantly higher perceived engagement. They felt the interaction helped them focus on relevant content ($M=5.75/7$) and increased active involvement ($M=6.13/7$).
• Workload: Contrary to the fear that interaction adds cognitive load, the NASA-TLX scores showed the interactive condition had slightly lower overall workload ($M=49.8$ vs. $52.8$). Participants felt less frustrated and exerted less mental effort, likely because the interaction structured their attention and pacing.
• Behavioral Changes: Participants asked fewer questions in the interactive condition ($M=3.75$) compared to the non-interactive condition ($M=4.38$), suggesting the interaction helped them self-solve or understand the material more efficiently without needing conversational clarification.
• Usability: The system received high usability scores (SUS $M=81.6$).

5. Significance and Future Work

• Design Implications: The paper demonstrates that interaction design is a critical variable in AI education, not just a UI feature. It suggests that "lightweight" interactions can effectively bridge the gap between passive AI consumption and active learning.
• Scalability of Design: The think-aloud session revealed that different pedagogical goals require different interaction modalities (e.g., "scratch-off" for hints, "reordering" for instructions, "progressive reveal" for modeling).
• Limitations & Future Directions:
  • The study is preliminary (small N, homogeneous population).
  • The exact cognitive mechanism is unclear: Did the interaction improve learning by fostering constructive engagement, or simply by slowing down reading to ensure attention?
  • Future work requires larger-scale studies, diverse learner populations, and richer behavioral data (e.g., eye-tracking, log analysis) to validate the robustness of these effects.

Conclusion: The paper successfully argues that integrating lightweight, deliberate interactions into LLM scaffolding can transform the learning experience from passive consumption to active engagement, offering a viable path forward for more effective AI-driven educational tools.