Towards Indo-Bloom: A Preliminary Study on Controllable Bloom's Taxonomy-aligned Question Generation in Indonesian

Automatic Question Generation (AQG) has the potential to revolutionize personalized education. However, ensuring that the questions generated are appropriate for a specific cognitive level, such as Bloom's Taxonomy, remains a major challenge, especially for languages with limited resources such as Indonesian. Current models often lack pedagogical control, failing to distinguish between surface information retrieval and deep conceptual understanding. Methodology: This article introduces the initial development of Indo-Bloom, a new framework and corpus designed for Bloom Taxonomy-aligned and controlled AQG in Indonesian. We present the Indo-Bloom Silver Corpus, consisting of 2,768 heuristically labeled question-answer-context triplets derived from three Indonesian QA sources, with a perfectly balanced cognitive distribution of 50% C1 (Remembering) and 50% C2 (Understanding). This corpus serves as the Silver Standard foundation (v0.1), pending expert validation for the Gold Standard release. To establish a baseline, we conducted a zeroshot evaluation on 50 expert-verified samples (25 C1-Remembering and 25 C2-Understanding) using state-of-the-art large language models (LLMs). Findings: The results show a significant performance gap between cognitive levels. While baseline models achieved high accuracy of 84% for C1 (Remembering), performance plummeted to 44% for C2 (Understanding). Error analysis revealed strong literal matching bias, where models misclassified C2 questions as C1 whenever the answer was explicitly stated in the source text. Contribution: This study provides empirical evidence that general LLMs are not pedagogically consistent for Indonesian AQG. Our main contribution is the Indo-Bloom Corpus, which serves as a foundation for refining encoder-decoder models (e.g., idT5) to achieve more precise cognitive control.