Fair Grades or Flawed Models? Evaluating the Trustworthiness of LLMs in Educational Settings

Read full text Summarise with VIRAT
Authors: Manar Hamed, Mariem Abdou, Heba Kamal Aslan, Tamer Arafa, Ghada Ahmed Khoriba
Cited by: 0
References: 21

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Large Language Models (LLMs) have recently attracted attention as they have become integrated into everyday applications.One essential use case in educational environments is their use as automated essay graders.However, this raises concerns about these models' reliability and susceptibility to adversarial manipulation.Therefore, this paper examines the robustness of six LLMs (GPT-4o, LLaMA3.3,Qwen2, Mixtral-8, Yi-Large, and Gemma3) against five attacks aimed at manipulating essay content.Our results show that the Qwen2 model was the most resilient, while the remaining LLMs were more vulnerable to several attack types.Two attacks, namely lexical obfuscation and semantic-preserving paraphrasing, led to substantial reductions in the average ±0.5 grading accuracy from approximately 0.69 in the clean setting to as low as 0.35, with effect sizes exceeding Cohen's d = 2 in the most severe cases.The study highlights the importance of developing robust defensive strategies to counter weaknesses in LLMbased educational testing systems.

Fields of Study & Subjects

Adversarial Robustness in Machine Learning Topic Modeling Artificial Intelligence in Education

Publication details

Journal / source
International journal of intelligent engineering and systems
Volume
19
Issue
4
Pages
334–350
Publication date
23 March 2026
Language
EN