The IDEA Framework for Metacognitively Regulated GenAI Use in Higher Education: Development and Exploratory Pilot Evidence

Synthesis: Wang et al. (2026) address the risk of "performance without learning" in Generative AI use by proposing the IDEA framework — Intent, Deconstruction, Expression, and Adaptation — a theory-informed metacognitive scaffold that embeds prompting within a cycle of planning, monitoring and AI Regulation in Education rather than treating it as an isolated technical skill. In an exploratory quasi-experimental pilot with 42 undergraduates (21 IDEA, 21 Prompt Engineering), IDEA-trained students produced substantially higher-quality prompts and final AI outputs than peers given structured Role–Task–Context–Format prompt-engineering instruction across all five task categories, with interaction logs showing observable enactment of the framework's core activities. On unaided tasks five days later the IDEA group retained advantages on selected tasks. The framework reframes GenAI use from passive content outsourcing into deliberate, evaluative, learner-regulated interaction.

Key Findings

The IDEA scaffold. The framework operationalizes metacognitively regulated GenAI use through four interdependent moves: articulate goals, audience, output format and constraints (Intent); break the task into modular sub-tasks and plan an interaction pathway (Deconstruction); communicate requirements through a structured prompt (Expression); and evaluate the output against task criteria and diagnose which phase to revise (Adaptation). Prompting is thereby nested inside a wider self-regulatory cycle, extending Self-Regulated Learning into the GenAI era. Table 1 in the paper provides a concrete instructional checklist (e.g. assign a role/persona, specify format constraints, diagnose whether a shortcoming stems from Intent, Deconstruction, Expression or the system).

Design. A quasi-experimental pilot compared IDEA-based instruction (n = 21) with structured prompt-engineering instruction (n = 21) based on the Role–Task–Context–Format approach. All participants used Tencent's Hunyuan Large Language Models (LLMs) (Hy3 Preview) through the Yuanbao interface, with the same model version and environment across conditions. Baseline and immediate post-instruction assessments each covered five AI-assisted tasks: lesson-plan design, cultural-activity design, behavioral-experiment design, literature review, and Writing. Fully adjusted regression models included instructional condition, baseline score, and prior Yuanbao-client use as covariates. IDEA is best understood as an iterative process — Adaptation may send a learner back to Intent, Deconstruction or Expression — not a one-pass linear sequence.

Results — prompt quality. IDEA-trained students produced higher-quality prompts in all five tasks. Adjusted differences favoring IDEA ranged from +11.77 points (literature review, p = .032) to +29.19 points (behavioral-experiment design, p < .001); all effects remained statistically distinguishable after Benjamini–Hochberg false-discovery-rate adjustment (qs ≤ .032).

Results — AI-output quality. Generated outputs based on post-instruction prompts also scored higher in the IDEA condition across all five tasks: +3.24 (lesson-plan, p = .008), +6.71 (behavioral-experiment, p = .027), +4.93 (cultural-activity, p = .002), +5.91 (literature review, p = .006), and +9.46 (academic writing, p < .001), all FDR-significant (qs ≤ .027).

Transfer. On unaided tasks completed five days later, the IDEA group demonstrated advantages on selected tasks, suggesting partial transfer beyond the scaffolded setting — though the authors caution that longer-term retention, transfer and Self-Regulated Learning change require direct measurement.

Significance and limits. The paper positions IDEA as a practical antidote to Cognitive Offloading and Over-Reliance: by requiring learners to plan, monitor and regulate their AI use, it aims to preserve Learner Agency and deep learning while capturing GenAI's efficiency. It complements AI Literacy interventions by offering a teachable Metacognition procedure rather than a content checklist, but the authors stress it is not a comprehensive AI-literacy framework and cannot substitute for disciplinary knowledge — domain expertise is especially consequential during Adaptation, since judging the plausibility and adequacy of an output requires it. The pilot is small, exploratory, single-site and non-randomized; the authors call for replication across disciplinary contexts and for longer-term learning-outcome measurement.

What this means for practice

• Instructors. Teach prompting as a four-phase regulatory cycle — Intent, Deconstruction, Expression, Adaptation — rather than a one-pass formula, since IDEA-trained students produced higher-quality prompts in all five task categories (adjusted gains of +11.77 to +29.19 points, all FDR-significant).
• Learners. Plan before prompting: state the goal, audience and constraints, break the task into sub-tasks, then evaluate each output against criteria before requesting revisions.
• Instructors. Assign an unaided transfer task after instruction; the IDEA group retained advantages on selected tasks five days later, but longer-term retention requires direct measurement.
• Designers. Embed revision loops in tools so a learner who sees a weak output is routed back to Intent, Deconstruction or Expression rather than nudged to re-prompt blindly.

Limitations

• An exploratory quasi-experimental pilot with 42 undergraduates at a single site, 21 per condition, and no randomization, so the IDEA advantage may reflect pre-existing differences rather than the scaffold.
• The framework cannot substitute for disciplinary knowledge: judging output plausibility during Adaptation depends on domain expertise learners may lack, and the authors state IDEA is not a comprehensive AI Literacy framework.
• The measures were task-specific prompt and output scores, not conceptual understanding, retention or independent problem solving, so claims about learning transfer remain unverified.
• Transfer evidence rests on unaided tasks completed only five days after instruction, with the authors calling for longer-term and cross-disciplinary replication.

Citation

Wang, X., Zheng, Z., Zhang, J., Hou, X., & Zhu, Z. (2026). The IDEA Framework for Metacognitively Regulated GenAI Use in Higher Education: Development and Exploratory Pilot Evidence. Computers and Education: Artificial Intelligence, 100657. https://doi.org/10.1016/j.caeai.2026.100657