Aouadi, Nour (2026) Can AI Solve It? Systematic Investigation of the Ability of Generative AI to Answer Database Exams. Masters thesis, Universität Rostock.
![[img]](https://eprints.dbis.informatik.uni-rostock.de/style/images/fileicons/text.png) |
Text
Nour_Aouadi__Master_Thesis.pdf Download (9MB) |
Abstract
Large Language Models (LLMs) such as Claude, Gemini, Kimi, and ChatGPT are increasingly used in higher education to support learning and solve academic tasks. This thesis investigates how reliably LLMs answer database examination questions and how their performance changes under different evaluation conditions. The study combines three empirical parts. First, Claude, Gemini, and Kimi were benchmarked on real Database and Data Warehouses exam questions using the professor’s grading criteria. Claude achieved the highest overall correctness with 85.6 %, followed by Kimi with 63.4% and Gemini with 59.4%. Second, selected questions were reformulated to test model robustness. The reformulated benchmark showed that stricter wording, output constraints, and multi-step requirements reduced answer quality, especially for Kimi and Gemini. Third, a real-life experiment showed that LLM assistance improved average correctness from 34.7% to 63.8 %, but the benefit depended on prompting strategy, context, input format, and the student’s ability to verify the generated answer. Overall, the results show that LLMs can solve many database-related exam tasks, but they are not consistently reliable. Their performance depends strongly on task design, prompt formulation, input modality, language, available context, and user interaction. This thesis therefore provides empirical evidence for the need to design AI-aware assessment methods in higher education.
| Item Type: | Thesis (Masters) |
|---|---|
| Subjects: | Autorenart > Studentische Arbeiten > Masterarbeit Autorenart > Studentische Arbeiten |
| Depositing User: | Dbis Admin |
| Date Deposited: | 08 Sep 2026 08:21 |
| Last Modified: | 08 Sep 2026 08:21 |
| URI: | https://eprints.dbis.informatik.uni-rostock.de/id/eprint/1158 |
Actions (login required)
 |
View Item |
