| C. Troussas, C. Papakostas, A. Krouska, Ph. Mylonas, C. Sgouropoulou |
| Personalized Automated Test Generation in Programming Education Using Fuzzy-Weighted Large Language Models |
| International Journal of Intelligent Systems, Wiley, September 2026 |
|
ABSTRACT
|
| The rapid adoption of large language models (LLMs) has stimulated research into automated assessment generation, yet most existing approaches either rely on generic prompts that overlook learner diversity or apply traditional personalization methods detached from generative AI. This creates a compelling need for frameworks that combine the linguistic flexibility of LLMs with principled learner modeling to deliver assessments that are both adaptive and psychometrically reliable. This paper introduces a generative-adaptive assessment paradigm that integrates fuzzy reasoning with large language models to address the content-exhaustion limitations of traditional item-bank-based adaptive testing systems. Learner proficiency is first estimated through a curriculum-aligned diagnostic assessment and normalized to the interval [0,1]. Trapezoidal membership functions, elicited from domain experts, transform these scores into fuzzy weights representing low, medium, and high proficiency. These weights are then injected directly into the ChatGPT prompt via the OpenAI application programming interface (API), shaping the target distribution of item difficulty during generation. The system architecture connects a learner interface, a fuzzy-weighted personalization layer, the LLM, and an item bank with automated validation filters, ensuring that generated assessments are coherent, balanced, and auditable. The contribution of this work lies in the operational integration of fuzzy learner modeling with LLM-based item generation through an explicit and auditable mapping from proficiency estimates to target difficulty distributions. The framework brings these components together within an end-to-end assessment pipeline that combines generation, personalization, validation, and empirical evaluation. Analysis of data from a controlled classroom evaluation involving 282 undergraduate students across multiple course sections demonstrated that the proposed approach significantly improved learning performance, assessment reliability, and perceived fairness compared to both manual assessment design and non-personalized LLM-based generation.
|
| 16 September, 2026 |
| C. Troussas, C. Papakostas, A. Krouska, Ph. Mylonas, C. Sgouropoulou, "Personalized Automated Test Generation in Programming Education Using Fuzzy-Weighted Large Language Models", International Journal of Intelligent Systems, Wiley, September 2026 |
[
BibTex] [
Print] [
 Back] |
