Scholarlink

This study investigates the potential of Large Language Models to grade geography homework worksheets from Secondary 1 students in Singapore, comparing their performance with that of a human teacher. Using a set of 12 structured questions and an explicit marking scheme, we evaluated two prompting strategies—lenient and strict—on OpenAI’s GPT-4o and Meta’s LLaMA-3-70B. GPT-4o achieved a 60% exact match rate with teacher-assigned scores and showed high reproducibility with minimal variance across runs. The models performed best on fact-based recall questions but struggled with context-dependent tasks involving textbook figures and diagrams. Teacher reviews found LLM-generated reasoning to be generally sound, though occasionally limited by prompt context or rigid interpretations. Our findings suggest that LLMs, when paired with precise prompts and human oversight, can assist teachers by automating routine grading tasks, especially for questions with well-defined evaluation criteria. Despite some limitations in generalizability and domain nuance, this can help reduce workload and enable more personalized student feedback, especially for practice tests.

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

LLM-Based Scoring of Secondary School Geography Worksheet: A Comparative Study of Accuracy, Reasoning and Reproducibility

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Divya Venkatraman,
  • Xin Ying Wong,
  • Jay Shah,
  • Vidyaraman Sankaranarayanan

摘要

This study investigates the potential of Large Language Models to grade geography homework worksheets from Secondary 1 students in Singapore, comparing their performance with that of a human teacher. Using a set of 12 structured questions and an explicit marking scheme, we evaluated two prompting strategies—lenient and strict—on OpenAI’s GPT-4o and Meta’s LLaMA-3-70B. GPT-4o achieved a 60% exact match rate with teacher-assigned scores and showed high reproducibility with minimal variance across runs. The models performed best on fact-based recall questions but struggled with context-dependent tasks involving textbook figures and diagrams. Teacher reviews found LLM-generated reasoning to be generally sound, though occasionally limited by prompt context or rigid interpretations. Our findings suggest that LLMs, when paired with precise prompts and human oversight, can assist teachers by automating routine grading tasks, especially for questions with well-defined evaluation criteria. Despite some limitations in generalizability and domain nuance, this can help reduce workload and enable more personalized student feedback, especially for practice tests.