Beyond Fluency: Cognitive Control and Pedagogical Quality in LLM-Based Automatic Question Generation — A Systematic Review of Bloom-Aligned Assessment Practices (2020–2025)

Aim/Purpose: This systematic review synthesises 69 primary studies (2020–2025) to map how contemporary automatic question generation (AQG) research operationalises cognitive control, quality validation, and language inclusivity, and to identify where the field still needs to make progress. Background: While large language models (LLMs) have resolved the fluency challenge in automatic question generation (AQG), they have not resolved the harder pedagogical challenge. The field must now shift from “fluent question generation” to “pedagogically valid question generation,” creating assessments that strictly align with cognitive frameworks such as Bloom’s Taxonomy and possess verifiable docimological quality. Methodology: A systematic literature review (SLR) was conducted across Scopus, IEEE Xplore, and SpringerLink, filtering 775 initial records to 69 primary studies (2020–2025) using PRISMA-guided inclusion/exclusion and quality assessment criteria. Contribution: Unlike prior AQG reviews that address technical performance or single pedagogical dimensions in isolation, this review is the first to systematically integrate three concerns: (1) Bloom-aligned cognitive control, (2) pedagogical validation beyond automated metrics, and (3) language inclusivity. By synthesising these pillars, the study provides a taxonomy of control methods, establishes expert evaluation as the validation reference standard, highlights the low-resource language gap (e.g., Indonesian), and proposes a Pedagogical Alignment Framework. Findings: The review reveals a clear architectural divide between prompt-based flexibility and fine-tuning precision. The data show that 88.4% of the studies addressed cognitive control and that all 69 studies (100%) applied some form of quality validation beyond n-gram metrics. However, only 24.6% did so with full docimological rigour. In addition, 87.0% of the studies focused exclusively on English-language contexts. The review also establishes that automated metrics such as BLEU and ROUGE are inadequate for measuring pedagogical quality. Recommendations for Practitioners: To ensure docimological quality, practitioners must mandate subject matter expert (SME) validation. Furthermore, institutional deployment should strategically align the AI architecture with the learning goal: choosing parameter-efficient fine-tuning (PEFT) for strict cognitive compliance (e.g., high-stakes exams) or LLM prompting for flexible, exploratory tasks. Recommendation for Researchers: Researchers and developers are encouraged to focus on creating datasets for non-English and low-resource languages to address equity gaps in educational technology. Impact on Society: By addressing the cognitive and validation challenges in AQG, this review promotes the development of more reliable, pedagogically sound, and equitable AI-supported assessment tools for diverse learner populations. Future Research: Future research should focus on empirically validating the proposed Pedagogical Alignment Framework, developing automated assessors to relieve the validation bottleneck, and expanding AQG applications in culturally contextualised and low-resource languages.