From Content Generation to Learning Support: Pedagogy-Guided Generative Video Tutors for STEM Learning

Synthesis: PIVOT treats the instructional video as something to be verified and assessed rather than merely rendered. Its three stages wrap Generative AI output in a pedagogy-structured storyboard, a Multimodal AI verification harness, and misconception-aware remediation before a learner ever sees the result. Across 40 STEM topics spanning four domains, PIVOT scored 7.58 on a 10-point frame-level video rubric against 6.62 and 6.20 for two code-centric baselines, and 0.89 on narration quality. In expert evaluation, 96.9% of 32 instructors said the video → quiz → remediation workflow worked better than an instructional video alone, which is the paper's real claim: pedagogical structure, not rendering fidelity, is what makes generated video teach.

Key Findings

1. Pedagogy-guided planning beats content-only generation. PIVOT's storyboards are built from learning objectives, prerequisite activation, stepwise explanations, visual plans, worked examples and diagnostic probes, and its videos reached an overall 7.58 out of 10 against 6.62 (TheoremExplainAgent) and 6.20 (Code2Video).
2. Visual richness produced the largest margin. PIVOT scored 6.06 on richness versus 4.23 and 4.22 for the baselines, while its layout advantage was far smaller (9.10 against 9.00), so the framework's clearest automatic-evaluation win is coverage of instructional content, not tidiness.
3. Layout verification carries the layout score. Removing the layout-consistency module dropped layout accuracy from 9.10 to 7.87, with the authors attributing the framework's spatial clarity to that check rather than to the underlying video generator.
4. Factual and cognitive checks move narration only modestly. Without factual verification accuracy and depth fell 0.87 → 0.85 on a 0–1 scale, and without cognitive verification logical flow fell 0.91 → 0.89 — small but directional effects on explanation reliability and coherence.
5. Experts preferred the loop over the artifact. 96.9% of the 32 participating STEM instructors reported that the instructional-video → assessment → targeted-remediation sequence provided a more effective learning experience than watching an instructional video alone.
6. Gains concentrated where misconceptions are procedural. Mean expert ratings rose by +1.67 points in physics and +0.76 in computer science over the baseline across comparable dimensions, with smaller changes in mathematics and chemistry, whose rule-driven topics a single video can already convey.

What the verification harness actually verifies

The design insight worth taking from PIVOT is that generation quality and instructional quality are separate problems, and this paper addresses the second with three explicit checks. Layout consistency detects overlap, spacing problems and excessive information density; factual consistency checks formulas, symbolic transitions and algorithmic procedures; cognitive clarity checks logical progression and coherence of the explanations. Each check feeds an iterative low-temperature repair pass conditioned on execution traces, with GPT-5.2 as the primary model and Claude Sonnet 4.6 handling persistent failures. The ablation shows these components are not interchangeable: layout verification moves visual quality, while factual and cognitive verification move narration. That is a more useful design statement than a single aggregate score, because it tells builders which failure they are buying protection against.

Planning as an intermediate representation

Rather than prompting a video model directly from a topic string, PIVOT first emits an instructional storyboard and uses it as the semantic backbone for executable Manim code. The storyboard is what makes Scaffolding explicit, since each of its components maps to an established instructional function: objectives focus attention, prerequisite activation manages prior knowledge and reduces cognitive load, worked examples demonstrate procedural reasoning, and diagnostic probes create early opportunities to catch Misconceptions about AI. Shot-level structure also makes the pipeline auditable — each shot carries its own local goal, narration, visual description and duration constraint, so a verification failure can be localized and repaired instead of triggering regeneration of the whole lesson. The authors present this intermediate representation as the mechanism that keeps generated video aligned with instructional intent.

Assessment and remediation close the loop

The third stage connects the video to post-video formative assessment. Quiz items are derived from the storyboard but deliberately avoid reusing the video's examples, equations or numerical values, so a correct answer indicates conceptual transfer rather than recall of what was just shown. Every answer option receives feedback, and an incorrect response triggers a remediation video conditioned on the question, the chosen option, the learner's answer and the correct reasoning, contrasting the wrong reasoning with the right one. This contrastive framing is where the expert ratings were strongest: error explanation showed the largest improvement over the baseline, and the authors read that as evidence that misconception-aware remediation supports conceptual understanding in a way standalone instruction does not.

What this means for practice

• Instructors. Adopt the three-part sequence — instruction, then a diagnostic question, then a targeted explanation of the specific error — rather than the video alone; 96.9% of the expert panel judged that sequence more effective than a standalone video.
• Instructional designers. Specify the storyboard before the render: objectives, prerequisite activation, stepwise explanation, visual plan, worked example and assessment probe are the components the framework's quality checks are written against.
• Developers. Separate verification from generation. Layout, factual and cognitive checks move different quality dimensions in the ablation, so a single generic "quality" pass will leave at least one of them uncovered.
• Researchers. Treat interaction-level remediation as the weaker half of this design; the pipeline reacts to one response at a time and the authors name long-term learner modeling as the missing piece.

Limitations

• Adaptation is interaction-level only: the system conditions remediation on the immediate response and does not model long-term knowledge tracing, learner memory, curriculum progression or individual difficulty across extended sessions.
• The multi-stage pipeline — storyboard generation, executable rendering, verification and remediation — costs substantially more computation than direct multimedia synthesis, and the paper reports no efficiency or cost figures.
• Expert evaluation used a questionnaire with 32 instructors and shuffled presentation order, so it measures perceived quality and perceived learning effectiveness rather than measured learning outcomes.
• Generation depends on GPT-5.2 with Claude Sonnet 4.6 as fallback and Gemini-2.5-Flash as the automatic judge, so the reported scores are tied to those specific model versions.

Connected Concepts

• Scaffolding
• Misconceptions about AI
• Video in Education
• Generative AI
• Multimodal AI
• Formative Assessment
• Automated Question Generation
• Prior Knowledge
• Adaptive Learning
• Feedback
• Cognitive Psychology
• Transfer of Learning
• STEM Education
• Human AI Collaboration

Connected Articles

• ANVIL: Analogies and Videos for Lecturers — ANVIL: Analogies and Videos for Lecturers
• When Saying No Makes Better Videos: Designing Dual Gatekeeping for Pedagogically Grounded AI Content Creation — When Saying No Makes Better Videos: Designing Dual Gatekeeping for Pedagogically Grounded AI Content Creation
• Advancing diagram-based reasoning in AI tutoring systems: a structural approach for STEM education — Advancing diagram-based reasoning in AI tutoring systems: a structural approach for STEM education
• The Correct Answer Trap: Pedagogically-Grounded Detection and Feedback for Hidden Misconceptions — The Correct Answer Trap: Pedagogically-Grounded Detection and Feedback for Hidden Misconceptions
• Adaptive Scaffolding for Cognitive Engagement in an Intelligent Tutoring System — Adaptive Scaffolding for Cognitive Engagement in an Intelligent Tutoring System
• SAVVY: Student Attention Visualization for Video-based Learning Analysis — SAVVY: Student Attention Visualization for Video-based Learning Analysis
• AI-Guided Learning: Research on Knowledge and Skill Acquisition Support Methods Using Deep Learning Audio-Video Processing Techniques — AI-Guided Learning: Knowledge and Skill Acquisition Support Using Deep Learning Audio-Video Processing
• An Interpretable Closed-Loop Intelligent Tutoring System for Multimodal Affective Feedback in Asynchronous Presentation Training — An Interpretable Closed-Loop Intelligent Tutoring System for Multimodal Affective Feedback in Asynchronous Presentation Training

Citation

Ma, X., Wang, S., He, G., Zhang, Y., Wang, C., Lan, Y., & Qian, W. (2026). From Content Generation to Learning Support: Pedagogy-Guided Generative Video Tutors for STEM Learning. arXiv:2609.24083.