CourseBlueprint: A Structured Pipeline for Adaptive Pedagogical Video Generation Grounded in Course Corpora

Synthesis: Islam et al. (2026) address a core limitation of generative text-to-video for education: while visually fluent, such systems lack pedagogical content knowledge (PCK). CourseBlueprint provides a structured pipeline producing adaptive pedagogical videos grounded in a course corpus (undergraduate biomedical-imaging course BMED 2300, 23 lectures, 1,116 slides). The pipeline includes four components with typed intermediate representations and validation: a scaffolding module with prerequisite concept graphs, an adaptive controller assigning style specifications per learner persona, an engagement generator using a fixed rhetorical contract (hook→retrieval→core→analogy→forward contract), and a deterministic slide-image override mechanism. Ablation results show removing the engagement contract crashes the engagement score from 5.00 to 1.20. The slide override converts a 0/9 corpus-grounding failure into 9/10 successful matches. This work demonstrates that Generative AI for education needs explicit pedagogical structure — not just fluency — and connects to Scaffolding, Personalized Learning, and Generative AI research in Higher Education.

What this means for practice

• Designers. Ground generation in a real course corpus instead of the model's pedagogical content knowledge alone. The pipeline indexes twenty-three lectures and 1,116 slides from BMED 2300 so each retrieved unit links a textual explanation to a specific instructor slide image.
• Designers. Make engagement an explicit typed contract rather than an instruction to "be more engaging": removing the engagement contract crashed the engagement score from 5.00 to 1.20, while the hook→retrieval→core→analogy→forward structure held the score up.
• Designers. Add a deterministic override for slide reuse. When retrieval confidence is high, reusing the instructor slide converted a 0/9 corpus-eligible-slide failure into 9/10 successful matches on the same topic.
• Designers. Resolve prerequisites before generation. The scaffolding module builds a prerequisite concept graph, applies a cycle-break algorithm, and colors nodes by stage so the build chain (simple back-projection → blurring → filtering → filtered back-projection) is validated rather than left to the generator.
• Designers. Decide the accuracy-cost trade-off deliberately: the optional verifier is disabled in the reported experiments, so factual correctness is supported by retrieval but not independently verified, and enabling it is the cheaper fix than regenerating video.

Limitations

• The evaluation covers ten generated videos from five in-corpus topics and a single learner persona, so it cannot establish robustness across learner profiles or topic types; out-of-corpus topics fall back to placeholder slides and lose the course-grounding benefit entirely.
• No human learners were measured. The authors state this limits any claim about learning effectiveness and describe a planned within-subjects study with n=20 undergraduate biomedical-engineering students, Latin-square topic ordering, and 5-point Likert ratings per PCK dimension as the next step.
• The judge is drawn from the same model family as the generator, which raises self-preference risk; the reported within-judge variance is small (σ̄=0.063, α≈0.97), but that does not establish agreement across model families.
• The ablation isolates only the engagement module, with scaffolding-off and adaptive-off conditions left to future work, so the results are a focused contrast between the full system and a no-engagement variant rather than a causal decomposition of all four components.

Citation

Md Zabirul Islam, Md Motaleb Hossen Manik, Ge Wang (2026). CourseBlueprint: A Structured Pipeline for Adaptive Pedagogical Video Generation Grounded in Course Corpora.