Scholarlink

<p>The success of responsive science teaching to elicit, foreground, and pursue the substance of students’ ideas depends on how teachers can recognize and respond to these ideas and connect them to disciplinary perspectives. One approach to supporting responsive teaching is to design instruction that facilitates students’ scientific sensemaking and invites teachers to consider possible student ideas. Generating students’ ideas, however, is labor-intensive as it requires understanding of learning content and instructional contexts. In this work, we explore the capacity of several large language models (LLMs), including OpenAI’s GPT-5-mini, GPT-4o, Claude’s Sonnet 4, Gemini 2.5 Flash, Mistral 7B, and Llama-4-17B, to simulate 8820 students’ ideas in science learning activities across domains (life sciences, physics, and chemistry) and grade levels (elementary, middle, and high school) to support instruction planning. Findings indicate that the LLM-simulated responses are realistic and are mostly at the target grade levels for knowledge scope and readability. We observe variations in performance across LLMs and grade levels, with LLMs producing more within-scope ideas for high school lessons and overly complex ideas for lower grades. Interviews with six teachers using the LLM-generated ideas with their own lesson plans reveal that the simulated student responses are overall realistic and align with the lessons’ objectives. Responses are most useful when they spark teachers’ sensemaking about what students know and suggest relevant instructional scaffolds. We discuss how to improve LLMs’ simulations of student thinking and promote responsive science teaching with AI-generated insights.</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

Exploring the Capacity of Large Language Models to Simulate Students’ Scientific Thinking: Insights for Responsive Teaching

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月21日

查看全文
  • Ha Nguyen,
  • Jie Cao

摘要

The success of responsive science teaching to elicit, foreground, and pursue the substance of students’ ideas depends on how teachers can recognize and respond to these ideas and connect them to disciplinary perspectives. One approach to supporting responsive teaching is to design instruction that facilitates students’ scientific sensemaking and invites teachers to consider possible student ideas. Generating students’ ideas, however, is labor-intensive as it requires understanding of learning content and instructional contexts. In this work, we explore the capacity of several large language models (LLMs), including OpenAI’s GPT-5-mini, GPT-4o, Claude’s Sonnet 4, Gemini 2.5 Flash, Mistral 7B, and Llama-4-17B, to simulate 8820 students’ ideas in science learning activities across domains (life sciences, physics, and chemistry) and grade levels (elementary, middle, and high school) to support instruction planning. Findings indicate that the LLM-simulated responses are realistic and are mostly at the target grade levels for knowledge scope and readability. We observe variations in performance across LLMs and grade levels, with LLMs producing more within-scope ideas for high school lessons and overly complex ideas for lower grades. Interviews with six teachers using the LLM-generated ideas with their own lesson plans reveal that the simulated student responses are overall realistic and align with the lessons’ objectives. Responses are most useful when they spark teachers’ sensemaking about what students know and suggest relevant instructional scaffolds. We discuss how to improve LLMs’ simulations of student thinking and promote responsive science teaching with AI-generated insights.