CS MSc Thesis Presentation Day 11 June 2026

Thursday 11 June is a day for coordinated master thesis presentations in Computer Science at Lund University, Faculty of Engineering. Eleven theses will be presented.

You will find information about how to follow along under each presentation. The presentations will take place in E:2116 and E:4130 (Lucas). A preliminary schedule follows.

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.

Many more presentations will take place on 4 June, see this link, and some on other dates see more on this link.

09:15-10:00 in E:2116

• Presenters: Samuel Högfeldt, Melissa Engberg
• Title: Computationally limited product type classification
• Examiner: Elizabeth Bjarnason
• Supervisors: Noric Couderc (LTH), Andreas Björklund (Voyado)

Nowadays large language models are popular because of their abilities to help with a wide berth of different tasks. This has resulted in companies and organizations relying on large language model companies for their own ability to deliver products and services. The large language models require considerable amounts of computational power and are often closed source, hosted in external datacenters, requiring subscriptions to be used. This paper explores using open-source small language models to classify products within fashion based on text inputs. Their performances is measured and evaluated on hardware usage, existing metrics as well as custom metrics proposed in this paper.

Link to popular science summary to be added

10:15-11:00 in E:2116

• Presenters: Albin Nyström, Aron Somi
• Title: Iterative Tensorwise Quantization of Neural Networks Using MLIR
• Examiner: Flavius Gruian
• Supervisors: Noric Couderc (LTH), Felix Malmsjö (Inceptron)

As Large Language Models (LLMs) continue to grow in size, so do their memory and computational requirements. Quantization is a common method for reducing these requirements to improve efficiency and meet hardware specifications, though often at the cost of model quality. Existing quantization methods typically rely on advanced calibration algorithms and perform one-shot quantization. This thesis explores iterative tensor-wise mixed-precision quantization using Bayesian optimization for two different quantization methods. Our solution leverages a MLIR backend to decrease the cost of exploring a large set of quantization configurations. The results demonstrate both the feasibility and the challenges of applying Bayesian optimization to LLM quantization. In particular, the large search space makes it difficult to consistently identify optimal quantization configurations, indicating that additional methods for constraining or guiding the optimization process are needed.

Here is a link to the popular science summary

10:15-11:00 in E:4130 (Lucas)

• Presenters: Jiuen Feng, Ruiying Zhu
• Title: BEV-Grid vs. Query-Based Architectures for Camera-Only Free-Space and Occupancy Prediction across varying Feature-Space Sizes
• Examiner: Pierre Nugues
• Supervisors: Eren Aksoy (LTH), Angelo Delli Santi (Zenseact AB), Johan Jaxing (Zenseact AB)

This thesis investigates camera-only BEV representation learning for short-range free-space and occupancy estimation, focusing on how intermediate BEV representations affect dense spatial prediction when the feature space is compressed. We compare dense voxel-lifting and Transformer-based sparse-query models in terms of prediction quality, computational efficiency, and representation compactness.

We adapt a Simple-BEV-based dense model and a PETR/PETRv2-based sparse-query model for BEV free-space and occupancy prediction. Two BEV target definitions are used: an HD-map-based target derived from nuScenes and a binary occupancy target constructed from Occ3D. The models are evaluated by varying the intermediate feature-space size, allowing the effect of feature-space compression on BEV representation capacity to be analyzed. Lastly, the resulting trade-off between quality and efficiency is assessed using IoU-based segmentation metrics and runtime measurements.

Here is a link to the popular science summary

11:15-12:00 in E:2116

• Presenters: Nora Wirtén, Joel Dahl
• Title: Generating High Quality Visualization of Software Architecture using Multi-Agent AI
• Examiner: Ulf Asklund
• Supervisors: Emma Söderberg (LTH), Andreas Bexell (LTH), Örjan Percy (Bosch)

As software systems continue to grow in size and complexity, architecture diagrams become increasingly important. However, many tools that generate diagrams automatically produce diagrams that suffer from inconsistent annotation, inaccuracy, and excessive complexity. To make generated diagrams genuinely useful, the diagrams must be easy to read, correct, and tailored to the person who will use the diagram. Different tasks require different diagrams.

This thesis explores the use of multi-agent AI to create custom diagrams tailored to the user's needs while still being human-readable. A literature study sets a base for layout quality and an empirical study for what developers expect from diagrams. This base sets a framework for iterative prototype development and evaluation. The multi-agent system utilizes several agents with different roles, working together to create diagrams depending on context extracted through prompts. The final prototype generates diagram-as-code and drag-and-drop editable diagrams, and is evaluated through another empirical study.

Link to popular science summary to be added

11:15-12:00 in E:4130 (Lucas)

• Presenters: Hugo Orrberg, Olof Bengtsson
• Title: Leveraging Histopathology Images for Rejection Grading and Multi-modal Survival Prediction in Heart Transplantation
• Examiner: Jacek Malec
• Supervisors: Pierre Nugues (LTH), Johan Nilsson (Faculty of Medicine, LU), Henry Pigot (Faculty of Medicine, LU)

A critical challenge in post-heart transplant care is balancing the risks of over-treatment against organ rejection. This thesis evaluates an ML framework designed to predict long-term survival by analyzing biopsy WSIs and medical records. A model was first trained to recognize rejection (ACR) grades before being adapted for survival prediction. The results confirm a correlation between ACR grades and 5-year survival, demonstrating that pretraining on these grades is essential for model performance. The data shows that the image-based model, combined with tabular data, achieves a slight performance improvement over survival models based only on a pathologist’s ACR grade. These findings suggest that the inconsistent and discrete nature of the ACR scale is too coarse to capture all prognostic information present in the tissue. By utilizing raw images instead of grades, finer indicators of risk can be identified, laying the groundwork for more consistent and precise monitoring of heart transplant recipients.

Here is a link to the popular science summary

13:15-14:00 in E:2116

• Presenters: Olof Gilland, Kaspian Garpvall
• Title: Establishing Practices for Sharing ESS Control System Data with External Researchers
• Examiner: Per Runeson
• Supervisors: Fredrik Edman (LTH), Karin Rathsman (ESS), Timo Korhonen (ESS)

Large-scale research infrastructures generate operational data with significant potential for external research, but sharing such data is difficult when dat responsibility, sensitivity assessment, authorization, legal conditions, and prior sharing practices are unclear. This thesis investigates how control system data from the European Spallation Source can be shared with external researchers in a secure, useful, and repeatable way.

Using design science research, the study combines literature review, document analysis, exploratory and evaluative interviews with ESS stakeholders, solution design, and prototype development. The main contribution is a proposed data-sharing framework consisting of a role-based workflow, authorization scheme, sensitivity and shareability classification, release request form, and upload template. The findings show that repeatable sharing should begin with a layered process, clear authorization criteria, baseline sensitivity classification with dataset-level review, standardized documentation, and organizational policy support, rather than extensive restructuring.

Link to popular science summary to be added

13:15-14:00 in E:4130 (Lucas)

• Presenters: Märta Holmquist, Emma Kujala
• Title: Quantization Techniques for Memory-bound Transformers on Ethos-U Hardware Accelerators
• Examiner: Sven Robertz
• Supervisors: Flavius Gruian (LTH), Oscar Andersson (Arm), Adrian Lundell (Arm)

Efficient deployment of machine learning models on resource-constrained hardware remains a major challenge. In particular, edge devices running on microcontrollers or neural processing units require optimized models that balance accuracy, latency, and energy consumption. Quantization is a widely adopted technique for model compression and acceleration, where floating-point operations are approximated using lower-precision representations. Recent work has focused primarily on quantization strategies for Large Language Models (LLMs), while other transformer architectures have received comparatively less attention despite suffering from similar quantization challenges. This paper investigates the effects and challenges of applying state-of-the-art LLM-inspired quantization approaches to transformer encoder models, specifically Vision Transformers, targeting deployment on Arm Ethos-U85. The Arm Ethos-U series of microNPUs provides hardware acceleration for ML inference. While optimized for int8 quantization, newer quantization schemes and the use of sub-int8 formats remain largely unexplored, providing an opportunity to evaluate these approaches under realistic hardware constraints.

Here is a link to the popular science summary

14:15-15:00 in E:2116

• Presenters: Mattias Mc Mullin, Christoffer Fjällborg Rinaldo
• Title: Performance Analysis of JastAdd
• Examiner: Görel Hedin
• Supervisor: Christoph Reichenbach (LTH)

JastAdd is a framework for building compiler systems. It is used to implement the Java compiler ExtendJ. We investigate how we may determine its performance characteristics, and where time is spent during compilation. To answer these questions, we extend JastAdd with various instrumentation capabilities, and use these to benchmark the ExtendJ implementation. We determine that our benchmarking methodology is sound. We produce a dataset, including a detailed trace log, inclusive/exclusive timings, as well as timing distributions, of various ExtendJ components. From the data, we calculate aggregate statistics. We also derive a call graph, describing the execution flow during compilation, enriched with additional data. We implement tooling to explore this call graph interactively. We report the fraction of time utilised by each part of the ExtendJ compiler, on different levels of abstraction. Our methodology and dataset may be used to further improve ExtendJ and JastAdd in the future.

Here is a link to the popular science summary

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

• Presenters: Simon Ghidini, Alexander Jansson
• Title: Two-Layer Rate Controller for Video Encoding
• Examiner: Per Andersson
• Supervisors: Flavius Gruian (LTH), Viktor Edpalm (Axis Communications)

Video encoding systems rely on rate control (RC) algorithms to regulate bit rate while maintaining visual quality. This thesis investigates a two-layer RC architecture consisting of a slow control layer which produces candidate results and a fast decision layer which selects a final answer. Multiple prototype configurations were implemented and evaluated against a reference RC using bit rate analysis together with PSNR, SSIM, and VMAF. The results show that the proposed architecture achieves competitive perceptual quality compared to the reference implementation, with some configurations outperforming the reference in VMAF. 
However, several prototypes exhibit weaker bit rate regulation and exceed the target bit rate for extended periods. The results indicate that separating rate control into layers with different characteristics can improve perceptual quality at the cost of bit rate stability. The results also found prototypes that outperform the reference and other prototypes under certain testing conditions.

Here is a link to the popular science summary

15:15-16:00 in E:2116

• Presenters: Jonatan Svahn, Jacob Johansson
• Title: Improving the Compilation Time of Semantically Similar Cypher Queries
• Examiner: Görel Hedin
• Supervisors: Niklas Fors (LTH), Filip Hedén (Neo4j), Henrik Nyman (Neo4j)

Graph Database Management Systems (GDBMS) have grown rapidly in popularity over the years as a solution for managing graph-structured data. Neo4j is the most widely adopted GDBMS, and it uses a query language called Cypher. In this thesis, we present the implementation and evaluation for two new optimizations made to the Cypher compiler. The optimizations focus on semantically similar queries, which are queries that are syntactically different, but functionality-wise similar. The first optimization focuses on variable anonymization, which improves the ability to cache queries that only differ in terms of variable names. The second optimization focuses on adding a cache for parts of queries that can later be reused within or between multiple queries. The results show that in specific cases related to semantically similar queries, we gain a minor to significant improvement in compile time. The results also show a decrease in the code cache usage in the JVM for the second optimization in certain cases. When running more general query workloads, results show that our optimizations add a negligible amount of overhead.

Here is a link to the popular science summary

15:15-16:00 in E:4130 (Lucas)

• Presenter: Hongyu Shen
• Title: Fine-Grained Communication-Computation Overlap for MoE Models via Dual-Stream Pipelining
• Examiner: Michael Doggett
• Supervisors: Arseni Ivanov (LTH), Minyu Cui (Chalmers University of Technology)

Mixture-of-Experts (MoE) models offer high model capacity but suffer from severe communication overheads. In native PyTorch implementations, dispatch communication, computation, and combine communication are executed strictly in serial, leading to significant GPU idling and low hardware efficiency. This paper proposes a fine-grained communication-computation overlap mechanism for MoE models via dual-stream pipelining. By chunking tokens and leveraging a computation stream alongside a communication stream, we establish a three-stage pipeline. Experimental results demonstrate that our fine-grained pipeline significantly mitigates communication bottlenecks.

Link to popular science summary to be added