Classroom AI: Large Language Models as Grade-Specific Teachers

TL;DR

This work introduces a framework for finetuning LLMs to generate age-appropriate educational content across six grade levels, from lower elementary to adult education, and successfully adapts explanations to match students' prehension capacities without sacrificing factual correctness.

Abstract

Large Language Models (LLMs) offer a promising solution to complement traditional teaching and address global teacher shortages that affect hundreds of millions of children, but they fail to provide grade-appropriate responses for students at different educational levels. We introduce a framework for finetuning LLMs to generate age-appropriate educational content across six grade levels, from lower elementary to adult education. Our framework successfully adapts explanations to match students' comprehension capacities without sacrificing factual correctness. This approach integrates seven established readability metrics through a clustering method and builds a comprehensive dataset for grade-specific content generation. Evaluations across multiple datasets with 208 human participants demonstrate substantial improvements in grade-level alignment, achieving a 35.64 percentage point increase compared to prompt-based methods while maintaining response accuracy. AI-assisted learning tailored to different grade levels has the potential to advance educational engagement and equity.

Figures (11)

• Figure 1: (A) Overall framework of our approach. (B) Exemplar responses across different models.
• Figure 2: Results across evaluation criteria: (A) compatibility through integrated measure and ARI; (B) compatibility for each metric; (C) accuracy; (D) diversity gain and perplexity; and (E) survey results for type 1 questions.
• Figure 3: Survey results on Type 2 questions: (A) $\mathcal{D}_{NQ}$ and (B) $\mathcal{D}_{SQ2}$. Box plots show Q1 (question difficulty), Q2 (answer comprehensibility), and Q3 (model accuracy) across grade levels, with mean values as red dots on a five-point scale. Higher Q1 and Q2 results indicate lower question and answer difficulty; higher Q3 results indicate stronger accuracy. While $\mathcal{D}_{NQ}$ questions appear difficult for lower grades (low Q1 in A), answers remain comprehensible (high Q2 in A). Answer comprehensibility increases (higher Q2 in B vs. A) for grade-specific questions in $\mathcal{D}_{SQ2}$.
• Figure 4: Logit-lens visualization for the lower-elementary (top) and adult (bottom) models of LLaMA3.1:8B on the prompt, "Why is the sky blue? The sky is blue because". The bottom row for each figure shows the final output tokens, and each row above represents the top prediction at each transformer layer. Warmer colors (e.g., red) indicate higher confidence.
• Figure 5: Fields and subjects of questions used for finetuning.