Scholarlink

Deep learning based knowledge tracing (DLKT) models involve using deep learning techniques and learners’ exercise-answer data to model their knowledge state and predict their performance on future exercises. Explainable artificial intelligence (xAI) reveals that DLKT models rely on learners’ correctness, skills, and recency information from previously answered records. Inspired by the explanations, this study fine-tuned two large language models (LLMs) to perform the KT task by incorporating these key factors into the prompt. Experimental results demonstrate that Llama-3.2-3B achieves higher accuracy and precision scores compared to both Deep Knowledge Tracing (DKT) and Mistral-7B-v0.3. However, DKT outperforms both LLMs in terms of AUC, F1, and recall scores. Our approach, as a pilot study, highlights a promising direction for future research on the application of LLMs in KT.

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

Fine-Tuning Large Language Models for Knowledge Tracing Harnessing Insights from Explainable AI

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Deliang Wang,
  • Gaowei Chen,
  • Yu Lu

摘要

Deep learning based knowledge tracing (DLKT) models involve using deep learning techniques and learners’ exercise-answer data to model their knowledge state and predict their performance on future exercises. Explainable artificial intelligence (xAI) reveals that DLKT models rely on learners’ correctness, skills, and recency information from previously answered records. Inspired by the explanations, this study fine-tuned two large language models (LLMs) to perform the KT task by incorporating these key factors into the prompt. Experimental results demonstrate that Llama-3.2-3B achieves higher accuracy and precision scores compared to both Deep Knowledge Tracing (DKT) and Mistral-7B-v0.3. However, DKT outperforms both LLMs in terms of AUC, F1, and recall scores. Our approach, as a pilot study, highlights a promising direction for future research on the application of LLMs in KT.