AI Tutor Framework Uses Question Metadata to Shape Hints, Not Answers

A new framework for educational artificial intelligence does not try to answer reading-comprehension questions for students. Instead, it uses the hidden structure of each question—what kind of reasoning it demands, whether the answer is stated outright or must be inferred, and whether it concerns a single moment in a story or the whole narrative—to shape the hints an AI tutor produces. The approach, called AI as the More Knowledgeable Other (MKO) 2.0, was developed by researchers at Clemson University and Bournemouth University and published in Frontiers in Artificial Intelligence. Its central claim is deliberately modest: structured task information can measurably change how AI-generated guidance is worded, even though whether that guidance actually helps learners remains an open question.

The name borrows from sociocultural learning theory. In Vygotsky’s account of learning, a More Knowledgeable Other supports a learner in accomplishing tasks beyond their independent capability, ideally through contingent, minimally sufficient help rather than ready-made solutions. The authors are careful to say their system is MKO-inspired rather than a computational replica of a human tutor. It does not model what an individual learner knows, does not track performance history, and does not adapt to a person’s evolving understanding. It is task-sensitive, not learner-adaptive—a distinction the paper returns to repeatedly, because scaffolding theory holds that genuine instructional support depends on responding to learner performance, something this framework does not attempt.

The empirical foundation is the Quest Dataset, a publicly available collection of 6,989 question-answer pairs built for the Learning Agency Lab’s Quest for Quality Questions initiative and related to FairytaleQA, a narrative-comprehension resource derived from 278 children’s stories containing 10,580 expert-authored questions across seven narrative categories. Each Quest question carries three structured annotations: a primary comprehension attribute (action, causal relationship, character, feeling, outcome resolution, prediction, or setting), an evidence status flag separating explicit questions from those requiring implicit inference, and a scope flag distinguishing local questions from summary questions requiring integration across the text. The distribution is uneven: 91.1 percent of questions are local, 75.3 percent are explicit, and action (31.5 percent) and causal relationship (27.9 percent) dominate the attribute categories. Feeling questions stand out as the most frequently implicit category, with 382 of 665 requiring inference.

The first component of MKO 2.0 is a deterministic rule-based generator. It composes each scaffold from three parts: a core reasoning strategy tied to the question’s attribute, a modifier for evidence status, and a modifier for scope. A causal question prompts the learner to trace events that could explain an outcome; a prediction question prompts consideration of likely subsequent events; an implicit question is signaled as requiring clues to be combined; a summary question broadens attention across the narrative. Combining seven attributes with two evidence levels and two scope levels yields 28 distinct surface realizations, and the rule generator produced scaffolds for all 6,989 questions with no missing outputs. The authors stress that this completeness reflects implementation consistency—the mappings were defined by the rules themselves—not evidence of pedagogical value.

The second component is a controlled comparison of large language model behavior. The researchers drew an attribute-stratified sample of 140 questions, twenty from each of the seven attribute categories, using a fixed random seed. That sample contained 104 explicit and 36 implicit questions, and 121 local and 19 summary questions—naturally occurring proportions rather than a balanced design. Two LLM conditions were run on identical questions and narrative contexts using gpt-4.1-mini through the OpenAI API, with temperature set to 0 and a 120-token completion limit, one generation per item per condition. The generic condition received only the narrative text and the question, with instructions to provide one short hint that supports reasoning without stating the answer. The MKO-conditioned condition received the same input plus the three metadata fields, with instructions to use them in shaping the hint. No gold answer appeared in either prompt.

The paired comparison produced several statistically clear differences. MKO-conditioned scaffolds averaged 5.18 more words than their generic counterparts (median paired difference of 5.00), were longer in 115 of 140 pairs, and a Wilcoxon signed-rank test confirmed the length difference (W = 356.5, p < 0.001). Corpus-level lexical variation, measured with a 50-token moving-window type-token ratio, was slightly higher under MKO conditioning: 0.785 versus 0.765. The most striking difference concerned narrative-evidence cues—directions to consult the story, the text, or relevant clues—which appeared in 40.0 percent of MKO-conditioned scaffolds against 10.0 percent of generic ones, a pattern the exact McNemar test attributed overwhelmingly to cases where the cue appeared only under metadata conditioning (44 versus 2 discordant pairs, p < 0.001).

The evidence-status dimension showed the most targeted effect. Among the 36 implicit questions, inference-oriented language appeared in 13 MKO-conditioned scaffolds (36.1 percent) but in only one generic scaffold (2.8 percent), a significant paired difference (p = 0.002). Among the 104 explicit questions, such language was rare in both conditions and did not differ significantly. In other words, telling the model that a question requires inference changed the guidance it produced specifically for those questions. Yet other diagnostics were less favorable. Question-specific reference—whether the scaffold contains meaningful content words from the question itself—was already high in the generic condition (89.3 percent) and rose only to 93.6 percent under MKO conditioning, a non-significant difference. Supplying the narrative and question alone evidently enables substantial contextualization without any metadata.

Two results qualify the framework’s promise further. First, a lexical answer-overlap screening, which measures how many words from the reference answer appear in the scaffold, was essentially identical across the two LLM conditions (mean 0.091 generic versus 0.090 MKO-conditioned), meaning metadata conditioning did not reduce potentially answer-revealing language. The authors emphasize this is only a screening diagnostic: legitimate scaffolds may share vocabulary with answers, and semantically revealing hints may avoid exact overlap entirely. Second, a predefined global-scope diagnostic detected no scope-specific expressions in either condition among the 19 summary questions—a null result that may reflect weak scope-conditioned realization, limited sensitivity of the lexical pattern matching, or both, and the study cannot distinguish these explanations. A broad general reasoning-cue diagnostic also showed the opposite descriptive pattern, appearing in 76.4 percent of generic outputs versus 42.1 percent of MKO-conditioned ones, though the authors caution this broad measure says nothing about the quality or appropriateness of reasoning support.

The study’s limitations are stated plainly, and they matter for how the findings should be read. No learners, teachers, or independent expert raters were involved; the evaluation examined computational characteristics of generated text, not comprehension, engagement, or learning outcomes. The experiment used a single model configuration with one generation per item, and temperature 0 reduces but does not eliminate variability across repeated calls or model updates. The generic and MKO-conditioned prompts differed in both metadata availability and instructional framing, so observed differences cannot be attributed to metadata alone. Longer scaffolds, higher lexical variation, and more evidence-oriented language characterize different forms of output, but none of these demonstrates that students learn more from them. The authors also note that a meta-analysis of randomized preK-12 tutoring interventions reported an average effect size of 0.288 standard deviations, with stronger effects under sustained engagement and trained human support—evidence that high-quality instructional help works but is hard to scale through people alone, which motivates the entire enterprise.

What the study does establish is architectural feasibility. It shows that educational metadata, usually treated as descriptive labels for organizing datasets, can be repurposed as instructional-control signals that measurably shape what a language model says when it is asked not to answer but to guide. It also contributes a transparent, reproducible workflow: deterministic rules for the full dataset, a paired LLM design holding questions and narratives constant, and preserved prompts, code, and outputs. The authors position the contribution as a design space in which structured pedagogical representations constrain generative models rather than leaving instructional behavior implicit. Whether task-conditioned scaffolds improve reasoning, transfer, or help-seeking compared with generic guidance—and whether the framework extends beyond narrative comprehension to disciplinary reading or academic tutoring—remains, by their own account, a proposition for future research requiring direct human and learner-based evaluation.

Subject of Research: A metadata-driven framework that conditions AI-generated instructional scaffolds on the reasoning demands of narrative-comprehension questions.

Article Title: Artificial intelligence as the More Knowledgeable Other (MKO) 2.0: metadata-driven adaptive scaffolding for narrative comprehension

Article References: Akinwumi, P. O., Qian, M., Akande, I. O., & Babatope, O. A. (2026). Artificial intelligence as the More Knowledgeable Other (MKO) 2.0: metadata-driven adaptive scaffolding for narrative comprehension. Frontiers in Artificial Intelligence, 9, 1933178. https://doi.org/10.3389/frai.2026.1933178

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1933178

Keywords: artificial intelligence in education, large language models, instructional scaffolding, narrative comprehension, hint generation, educational NLP, reading comprehension, More Knowledgeable Other, Quest Dataset, metadata conditioning, tutoring systems, sociocultural learning theory