Scholarlink

Large language models (LLMs) have seen rapid advancements, which have made them widely accessible, with students increasingly relying on them for academic tasks. At the school level, many students use LLMs like ChatGPT and Gemini to complete essays and assignments, often copying generated content without engaging in critical thinking. This growing dependence raises concerns about declining creativity and originality in student writing. To address this challenge, our research introduces a robust classification framework capable of distinguishing between essays written by human and those generated by LLM. This paper presents a comparative analysis of two distinct approaches: (1) zero-shot prediction using the Google Gemini 2.0 Flash model and (2) a supervised Attention-Based Feed-Forward Neural Network (FFNN) trained on embeddings combined with linguistically motivated features. To assess the robustness of each method, a small subset of adversarial essays was added to the existing test dataset, simulating real-world scenarios. The dataset, sourced from Koike et al. (LLM-Generated Essay Detection Through In-context Learning with Adversarially Generated Examples, vol. 38. AAAI Press, pp. 21258–21266, 2024) [5], comprises student-authored essays and LLM-generated texts, providing a diverse foundation for training and evaluation. Our findings demonstrate that integrating advanced neural models with linguistic features significantly enhances classification accuracy. Additionally, we analyze the trade-offs between zero-shot prediction and supervised learning, outlining the impact and constraints of each approach. This study contributes to the ongoing effort to ensure academic integrity and promote originality in student writing.

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

Robust Classification of Human Written vs LLM Generated Essay Using Adversarial Training

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • R. Yogeshwaran,
  • R. S. Lysa Packiam,
  • Abirami Murugappan

摘要

Large language models (LLMs) have seen rapid advancements, which have made them widely accessible, with students increasingly relying on them for academic tasks. At the school level, many students use LLMs like ChatGPT and Gemini to complete essays and assignments, often copying generated content without engaging in critical thinking. This growing dependence raises concerns about declining creativity and originality in student writing. To address this challenge, our research introduces a robust classification framework capable of distinguishing between essays written by human and those generated by LLM. This paper presents a comparative analysis of two distinct approaches: (1) zero-shot prediction using the Google Gemini 2.0 Flash model and (2) a supervised Attention-Based Feed-Forward Neural Network (FFNN) trained on embeddings combined with linguistically motivated features. To assess the robustness of each method, a small subset of adversarial essays was added to the existing test dataset, simulating real-world scenarios. The dataset, sourced from Koike et al. (LLM-Generated Essay Detection Through In-context Learning with Adversarially Generated Examples, vol. 38. AAAI Press, pp. 21258–21266, 2024) [5], comprises student-authored essays and LLM-generated texts, providing a diverse foundation for training and evaluation. Our findings demonstrate that integrating advanced neural models with linguistic features significantly enhances classification accuracy. Additionally, we analyze the trade-offs between zero-shot prediction and supervised learning, outlining the impact and constraints of each approach. This study contributes to the ongoing effort to ensure academic integrity and promote originality in student writing.