CS MSc Thesis Presentations 28 May 2026

Thursday, 28 May there will be two master thesis presentations in Computer Science at Lund University, Faculty of Engineering.

The presentation will take place in E:1144 and E:4130 (Lucas).

Note to potential opponents: Register as an opponent to the presentation of your choice by sending an email to the examiner for that presentation (firstname [dot] lastname [at] cs [dot] lth [dot] se). Do not forget to specify the presentation you register for! Note that the number of opponents may be limited (often to two), so you might be forced to choose another presentation if you register too late. Registrations are individual, just as the oppositions are! More instructions for opponents are found here on the LTH thesis project page.

14:00-15:00 in E:1144

• Presenters: Yousif Kutaiba, David Larsson
• Title: A Comparative Evaluation of Hallucination Detection Methods for RAG Systems
• Examiner: Xuan-Son Vu
• Supervisor: Marcus Klang (LTH), Samir Jasarevic

Hallucination remains a central challenge for large language models, occurring when a model produces coherent but unsupported responses. This study defines hallucination as any response not entailed by a provided context, and evaluates several detection methods under this definition: specialized classifiers (LettuceDetect, HHEM 2.1, Lynx) and LLM-as-a-judge approaches (RAGAS, G-Eval), each tested with multiple underlying models. All methods are benchmarked on the RAGTruth dataset using AUROC, F1, MCC, latency, and cost. G-Eval with GPT-5.2 achieves the highest AUROC of 0.852 with strong cross-domain consistency, while LettuceDetect attains an AUROC of 0.809 at substantially lower latency and cost. A pipeline combining both methods is evaluated on a cybersecurity dataset, outperforming either method individually. The findings suggest that while LLM-as-a-judge methods deliver strong detection performance, encoder-based classifiers offer meaningful advantages in efficiency, and the choice of detection method matters as much as the underlying model.

Here is a link to the popular science summary

14:15-15:00 in E:4130 (Lucas)

• Presenters: Oscar Torstensson, Filip Greiff
• Title: AI-Based Profile Matching for Extruded Components
• Examiner: Jacek Malec
• Supervisor: Mathias Haage (LTH)

This thesis investigates how deep learning can be applied to enable efficient similarity-based retrieval of technical drawings in a large, unstructured industrial database. The work addresses challenges faced by engineers who spend significant time manually searching for existing designs or recreating them, even if they already exist. A system for image-based blueprint retrieval is developed using pretrained vision models, including Convolutional Neural Networks and Vision Transformers. The approach is evaluated using controlled benchmarks and real-world usability tests with domain experts. Results show that modern pretrained models can achieve high retrieval performance, especially when combined with preprocessing techniques, feature fusion and query area expansion. The system demonstrates robustness to variations in input, including handdrawn sketches. Still, we notice a gap in interpretation of similarity between human and model. This gap is explored, and we try to map it out and explain it. Furthermore, user studies indicate improved efficiency and workflow compared to traditional search methods. The findings highlight both the potential and limitations of applying deep learning to technical drawing retrieval in industrial contexts.

Here is a link to the popular science summary