Scholarlink

<p>As artificial intelligence systems integrate deeper into educational contexts, ensuring pedagogical alignment of large language models (LLMs) becomes critically important. This research investigates the alignment of GPT-4 with three established learning theories—Active Learning, Behaviorism, and Social Learning—to enhance explainability and addresses two questions: (1) how well GPT-4 outputs align with these learning theories, and (2) whether additional educational input improves this alignment. Two studies were conducted to answer these. Study 1 evaluated GPT-4’s alignment in tutoring via Khanmigo on Khan Academy courses by tagging instructional features in student prompts and outputs. The results revealed a strong alignment with Behaviorism (71.99%), but weaker alignments with Active Learning (24.63%) and Social Learning (10.74%). Study 2 assessed GPT-4’s performance as an instructional designer using a two-stage process to generate courses through the OpenAI API. Stage 1 generated benchmark courses with flexible inputs, and Stage 2 produced courses with theory-specific skill prompts. Pedagogical Alignment Distance (PAD) scores showed improved Behaviorism alignment in enriched courses (ΔPAD = 0.11); however, Active Learning (PAD = -0.17) and Social Learning (PAD = -0.29) alignments remained negative. The increased relevance of detailed input variables—particularly “skills” (0.35), compared to “course topic” (0.25) and “engagement methods” (0.29)—demonstrates GPT-4’s responsiveness to precise pedagogical cues. These results underscore the need for refined prompting strategies to enhance GPT-4’s alignment across diverse educational settings and advance explainable AI frameworks in education.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Explainable pedagogical alignment: auditing LLMs with learning theories

  • 期刊论文

  • 发表日期：

    2025年12月15日

查看全文
  • Daniel Domínguez Figaredo,
  • José Luis Fernández Vindel,
  • Laura Guerrero Puerta

摘要

As artificial intelligence systems integrate deeper into educational contexts, ensuring pedagogical alignment of large language models (LLMs) becomes critically important. This research investigates the alignment of GPT-4 with three established learning theories—Active Learning, Behaviorism, and Social Learning—to enhance explainability and addresses two questions: (1) how well GPT-4 outputs align with these learning theories, and (2) whether additional educational input improves this alignment. Two studies were conducted to answer these. Study 1 evaluated GPT-4’s alignment in tutoring via Khanmigo on Khan Academy courses by tagging instructional features in student prompts and outputs. The results revealed a strong alignment with Behaviorism (71.99%), but weaker alignments with Active Learning (24.63%) and Social Learning (10.74%). Study 2 assessed GPT-4’s performance as an instructional designer using a two-stage process to generate courses through the OpenAI API. Stage 1 generated benchmark courses with flexible inputs, and Stage 2 produced courses with theory-specific skill prompts. Pedagogical Alignment Distance (PAD) scores showed improved Behaviorism alignment in enriched courses (ΔPAD = 0.11); however, Active Learning (PAD = -0.17) and Social Learning (PAD = -0.29) alignments remained negative. The increased relevance of detailed input variables—particularly “skills” (0.35), compared to “course topic” (0.25) and “engagement methods” (0.29)—demonstrates GPT-4’s responsiveness to precise pedagogical cues. These results underscore the need for refined prompting strategies to enhance GPT-4’s alignment across diverse educational settings and advance explainable AI frameworks in education.