High-Order Question Generation: LLMs Move Beyond Bloom’s Taxonomy

Introduction

Good classroom questions do more than check whether students remember a fact. They can push learners to explain evidence, compare viewpoints, justify claims, or imagine alternative possibilities. That is why high-order questioning is closely tied to critical thinking. Yet designing such questions consistently is demanding for teachers, especially in multilingual education settings.

A paper accepted at LREC 2026, titled High-Order Question Generation in a Multilingual Educational Context, studies whether large language models can help with this task. Instead of focusing only on the widely used Bloom’s Taxonomy, the authors test two other prompting frameworks: Claim-Evidence-Reasoning and Divergent Questioning.

Key points

• The study broadens the framework. Much prior work on AI-generated educational questions has relied on Bloom’s Taxonomy to guide prompts. This paper asks whether other pedagogical frameworks can produce useful and varied high-order questions. Claim-Evidence-Reasoning encourages students to connect assertions with supporting information, while Divergent Questioning aims to elicit multiple possible answers or perspectives.

• The evaluation is multilingual. The study looks at Basque, Spanish, and English. This matters because educational AI is often tested primarily in English, even though real classrooms may involve languages with different levels of digital resources and model support.

• LLMs can generate questions, but quality is mixed. According to the abstract, both an open-source and a proprietary model generate questions rather effectively across all three languages. However, among answerable questions, only about half are recognized by teachers as high-order. This gap is important: a question may look complex on the surface without actually requiring deep reasoning.

• Alternative frameworks add diversity. A positive finding is that Claim-Evidence-Reasoning and Divergent Questioning lead to structurally and conceptually varied questions. Rather than replacing Bloom’s Taxonomy, these frameworks may complement it and expand the design space for AI-assisted teaching materials.

Why it matters

The paper highlights a central challenge for educational AI: generation is not the same as pedagogy. A model may produce fluent, plausible, and answerable questions, but that does not guarantee that the questions support analysis, evaluation, or creative reasoning.

For teachers, LLMs may still be useful as drafting tools. They can accelerate the creation of multilingual question sets and provide alternative phrasings or angles. But the study also reinforces the need for human review. Teachers remain essential for judging whether a question fits the lesson goal, the students’ level, and the kind of thinking the activity is meant to promote.

For developers of AI education tools, the findings suggest that systems should not be locked into a single educational taxonomy. Supporting multiple questioning frameworks could make generated materials richer and more adaptable. The multilingual dimension is also important: evaluation should account not only for language fluency, but also for instructional value across languages.

Overall, the study does not claim that LLMs have solved high-order question generation. Its contribution is more nuanced: language models can expand the teacher’s toolkit, but deciding what counts as meaningful high-order thinking still requires educational expertise.

Source: arXiv