Scholarlink

<p>The widespread application of generative AI in higher education has raised questions about how AI usage behaviors relate to students’ development of digital literacy. Existing research relies predominantly on correlation analysis, leaving potential causal pathways underexplored. This study proposes CD-LLM (Causal Discovery from LLM interaction logs), a hybrid causal discovery framework designed to explore potential causal pathways between AI usage behaviors and digital literacy from large-scale interaction logs. Experiments were conducted using 8,456 interaction records from the WildChat dataset, with digital literacy approximated using behavioral proxies aligned with the DigComp 2.2 framework. Results demonstrate that CD-LLM achieves an F1 score of 0.802 and a structural Hamming distance of 10.3, representing improvements of 12.0% and 23.7%, respectively, over the best baseline FGES. The framework proposes 21 data-consistent causal hypotheses among 16 variables, suggesting that interaction features may mediate relationships between prompt characteristics and digital literacy outcomes, with 85.7% of edges exceeding the 0.75 reliability threshold.</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

A causal discovery framework for analyzing the impact mechanism of generative AI on digital literacy: an LLM-based behavioral analysis approach

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月26日

查看全文
  • Rui Zhu,
  • Qialian Zou

摘要

The widespread application of generative AI in higher education has raised questions about how AI usage behaviors relate to students’ development of digital literacy. Existing research relies predominantly on correlation analysis, leaving potential causal pathways underexplored. This study proposes CD-LLM (Causal Discovery from LLM interaction logs), a hybrid causal discovery framework designed to explore potential causal pathways between AI usage behaviors and digital literacy from large-scale interaction logs. Experiments were conducted using 8,456 interaction records from the WildChat dataset, with digital literacy approximated using behavioral proxies aligned with the DigComp 2.2 framework. Results demonstrate that CD-LLM achieves an F1 score of 0.802 and a structural Hamming distance of 10.3, representing improvements of 12.0% and 23.7%, respectively, over the best baseline FGES. The framework proposes 21 data-consistent causal hypotheses among 16 variables, suggesting that interaction features may mediate relationships between prompt characteristics and digital literacy outcomes, with 85.7% of edges exceeding the 0.75 reliability threshold.