Feedback with Reasoning: Evaluating AI Systems That Explain, Scaffold, and Guide K-12 Learning

The ability of AI tutoring systems to provide not just correct answers but meaningful explanations — identifying student misconceptions, scaffolding productive struggle, and generating actionable guidance — represents one of the most consequential frontiers in educational technology. Our analysis of 183 papers in this category reveals a field undergoing a fundamental shift: from rule-based and template-driven feedback towards large language model (LLM)-powered systems capable of multi-turn dialogue, adaptive hint generation, and real-time pedagogical reasoning across subjects from elementary mathematics to high school essay writing.

The findings are both promising and sobering. On one hand, studies report meaningful learning gains — a 22.95% improvement in student outcomes with personalised feedback in one study, and an effect size of 0.56 for immediate correctness feedback in another. Comprehensive benchmarks such as TutorBench, MRBench, and MathDial now enable systematic evaluation of AI tutoring quality across multiple pedagogical dimensions. On the other hand, even leading models fall well short of expert tutoring: GPT-4 — the dominant commercial model at the time most of these studies were conducted — achieved only 56% on expert-curated tutoring tasks in TutorBench and was found to reveal solutions prematurely 66% of the time in dialogue-based tutoring scenarios. The field has made striking progress in measuring what AI tutors can do in the moment, yet remains critically weak in assessing what matters most: whether AI-generated feedback with reasoning builds durable understanding, supports metacognitive development, or inadvertently fosters dependency.

Perhaps most concerning for LMIC contexts, only 8 of the 183 papers examine non-English or cross-cultural feedback systems. The benchmarks, datasets, and evaluation frameworks that currently define quality in this space were overwhelmingly developed in high-income, English-language settings. This means the tools shaping global investment decisions about AI tutoring bear limited relevance to the linguistic, curricular, and infrastructural realities of the contexts where scalable, high-quality feedback is most urgently needed.