CS MSc Thesis Presentation Day 4 June 2026

Thursday 4 June is a day for coordinated master thesis presentations in Computer Science at Lund University, Faculty of Engineering. Ten 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 11 June, see this link, and some on other dates see more on this link.

09:15-10:00 in E:2116

• Presenters: Hanxuan Lin, Hanyu Lin
• Title: Multi-Stage Dashboard Generation over Graph Databases with LLMs
• Examiner: Emelie Engström
• Supervisors: Rushali Gupta (LTH), Niels De Jong (Neo4j), Noah Mayerhofer (Neo4j)

This thesis investigates natural language–driven dashboard generation over graph databases, with a focus on maintaining alignment between user intent and generated outputs. While prior work in natural language to visualization has improved single-chart generation, producing coherent multi-view dashboards remains challenging, particularly for graph-structured data where relationships and query constraints are central. To address this, the work frames dashboard generation as a multi-stage workflow rather than a single-step task. A pipeline is designed that separates intent interpretation, schema grounding, planning, query generation, and validation, with selective human-in-the-loop clarification. The proposed approach is implemented for a Neo4j database and evaluated against a baseline generation method through a comparative dashboard study and an interactive workflow study. Results suggest that the proposed workflow improves perceived dashboard quality and user experience, particularly in coherence, usability, and intent alignment. The findings indicate that structuring generation as an inspectable and iterative process is a promising direction for dashboard generation.

Here is a link to the popular science summary

09:00-10:00 in E:4130 (Lucas)

• Presenters: Wenrui Xie, Jiacheng Zhang
• Title: Diffusion Policy for Force- and Vision-Guided Robotic Liquid Pouring Task
• Examiner: Maj Stenmark
• Supervisor: Volker Krueger (LTH)

Liquid pouring is a high-precision, contact-rich task that remains challenging for imitation learning in laboratory automation and pharmaceutical manufacturing. This thesis extends diffusion policy for dual-arm robotic pouring by integrating force feedback with visual observations. Diffusion policy is suitable for pouring because it can model multi-modal action distributions and generate smooth temporal action chunks rather than isolated point-wise commands. To improve physical awareness, joint torque signals are added as proprioceptive inputs, enabling the policy to respond to load and interaction changes during execution. For safety, rice is used as a proxy medium instead of real liquid. The policy is trained on 150 human demonstrations and evaluated over 60 trials across three configurations: single-view baseline, multi-view baseline, and force-integrated multi-view policy. Results show that the force-integrated policy achieves the highest overall success rate and better task progression. Generalization to real liquids remains future work.

Link to popular science summary to be added

10:15-11:00 in E:2116

• Presenters: Julius Andreasson, Hussein Taher
• Title: Impact of JVM Optimization on Energy Consumption
• Examiner: Niklas Fors
• Supervisor: Christoph Reichenbach (LTH)

As the energy consumption of the IT industry grows, so does the need for sustainable computing. This thesis investigates the impact of JVM tiered compilation on energy consumption, specifically comparing Tier 1 and Tier 4.

The collection of measurement data during workloads, including power draw and JVM code execution, was automated using a containerized system. Using this system, we ran various benchmarks with different JVM configurations. Noise was minimized through locked CPU frequencies, disabled hyper-threading and kernel isolation. JVM code execution data was collected using DMCE instrumentation. We analyzed code heatmaps to correlate optimizations with energy-related impacts, increasing our understanding.

Our findings confirm that JVM configuration affects energy efficiency for workloads. We found that while power draw remains consistent between tiers, Tier 4 significantly reduces the energy consumption by minimizing execution time. The Kafka benchmark serves as an exception, where optimization overhead results in increased energy consumption.

Link to popular science summary to be added

11:15-12:00 in E:2116

• Presenters: Ludvig Lindholm, Danny Tang
• Title: Hierarchical Self-Supervised Learning for Semantic Embeddings of Crash Dumps
• Examiner: Niklas Fors
• Supervisors: Görel Hedin (LTH), Emil Eriksson (Schneider Electric Buildings AB), Jörgen Malmborg (Schneider Electric Buildings AB)

Analyzing software crash dumps to diagnose failures is currently manual and inefficient at Schneider Electric due to basic hashing methods. This thesis proposes a machine-learning embedding system that maps entire crash dumps into single semantic embeddings to automatically group them by underlying defects.

Since only a small fraction of the crash dumps are labeled, we opted for primarily using self-supervised training. To process these large files, we introduced a hierarchical encoder-based model that chunks data and combines it using learned, adaptive weighting. We also developed a novel attention-based loss function and a multi-phase curriculum to prevent overfitting.

Our self-supervised model significantly outperforms the ModernBERT baseline, nearly halving the false positive rate at a 95% retrieval threshold, proving its viability for clustering. A supplementary supervised model showed further promise in real-world manual testing despite ground truth data scarcity.

Link to popular science summary to be added

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

• Presenters: Haobo Zu, Yaliang Cai
• Title: Graph-Augmented RAG for Multimodal Document Question Answering
• Examiner: Xuan-Son Vu
• Supervisors: Pierre Nugues (LTH), Xuhao Zhang (Huawei Sweden R&D)

Large language models require grounded, structured evidence to answer questions reliably over private, multimodal enterprise documents. Existing graph-based RAG systems address retrieval structure but rarely combine multimodal document support, genuine multi-hop graph traversal, and adaptive retrieval control in a single deployable system. We present a graph-augmented RAG system built on RAG-Anything and deployed locally to preserve data privacy. We improve knowledge graph coherence through entity resolution, enable multi-hop reasoning via Personalized PageRank-based traversal combined with hybrid vector retrieval and cross-encoder reranking, and add an agentic pipeline that routes queries adaptively, detects retrieval insufficiency, and verifies answer grounding. The system is deployed as an interactive web application. Experiments on DocBench Academic, SurGE, and three multi-hop QA benchmarks show consistent improvements over the RAG-Anything baseline.

Here is a link to the popular science summary

13:15-14:00 in E:2116 N.B. No more opponents for this presentation

• Presenters: Antonio Krsoski, Emanuel Sjövall
• Title: Managing Failures in Stateful Systems
• Examiner: Alma Orucevic-Alagic
• Supervisors: Lars Bendix (LTH), Olof Englund (Neo4j)

Failures in stateful systems can cause persistent inconsistencies and are difficult to recover from. This is especially relevant in Kubernetes-based environments, where Operators automate the management of stateful applications but also introduce additional complexity and failure modes.

This thesis investigates failures in Operator-managed stateful systems within Neo4j’s cloud architecture. The initiating problem is that failures still occur despite existing testing and rollout strategies, and recovery often requires manual, non-standardized intervention. To address this, we examine two research questions: what root causes lead to failures in managed stateful systems, and how failures caused by Kubernetes Operators can be mitigated.

We identified 74 root causes and grouped them into 36 responsibility domains. Failures often stem from incorrect state evaluation and error-prone deployments. We developed Reconcile Dump to improve debugging and identified key operational pain points.

These findings suggest that improving reliability requires incremental improvements in observability, automation, and deployment design.

Here is a link to the popular science summary

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

• Presenters: Arvid Malm, Wolmar Boris-Möller
• Title: An On-Premise Diffusion Deep Research Model & Evaluation Framework for Resource-Constrained Environments
• Examiner: Pierre Nugues
• Supervisors: Patrik Edén (LTH), Thomas Drakengren (Qrendo AB)

State-of-the-art deep research models for generating long-form, citation-grounded reports typically rely on closed-source models and cloud infrastructure, limiting their use in high-security environments. This thesis explores whether such systems can be deployed fully on-premise for requirements management using a design science research approach with two main artifacts.

First, we develop an on-premise research agent, Diffusion On-premise Research Agent (DORA), inspired by test-time diffusion. Second, we design and implement an evaluation framework, Testing Local Deep Research (TLDR), that combines LLM-as-a-judge scoring with dynamic claim extraction and retrieval-based factual validation across four text quality metrics and correctness.

The results demonstrate that it is practically viable to run deep research architectures on-premise, granted appropriate design choices are made and smaller language models are used. The proposed model DORA outperforms all evaluated state-of-the-art models on the presented benchmark TLDR, employing a comparable state-of-the-art LLM.

Here is a link to the popular science summary

14:15-15:00 in E:2116

• Presenters: Emelie Tingberg, Victor Sannicolo
• Title: Test-Driven Development as a method for AI-generated code
• Examiner: Björn Regnell
• Supervisors: Per Runeson (LTH), Ivan Aladjoff (Decerno AB)

Large Language Models (LLMs) have created new possibilities for automating software development, but ensuring the quality of AI-generated code remains a challenge. This thesis investigates whether Test-Driven Development (TDD) can improve the quality of code generated by an LLM in a fully autonomous context at Decerno AB.

A TDD workflow was iteratively designed and compared against a non-TDD workflow using static code analysis. TDD reduced code complexity and duplication but introduced more quality issues overall, while non-TDD produced fewer but more severe issues. The analysis of generated tests in the TDD workflow revealed that the tests primarily focused on content verification rather than behavioral testing, contributing to a lower code coverage per test compared to non-TDD.

The conclusion is that TDD quality depends on which development priorities are valued. This suggests strict TDD may not be optimal in autonomous AI-driven contexts, motivating more flexible workflow adaptations.

Here is a link to the popular science summary

14:15-15:00 in E:4130 (Lucas) N.B. No more opponents for this presentation

• Presenter: Alve Lindell
• Title: GPU acceleration of LiDAR Inertial Odometry
• Examiner: Jonas Skeppstedt
• Supervisors: Michael Doggett (LTH)

LiDAR Inertial Odometry (LIO) provides highly accurate state estimation for autonomous robotic systems but suffers from significant computational bottlenecks, especially during the Generalized Iterated Closest Point (GICP) Point cloud registration. On resource-constrained edge devices, this high CPU demand can prevent real-time execution and constrain the available computational resources that could be required for other tasks. This thesis presents a GPU-accelerated implementation of Direct LiDAR-Inertial Odometry (DLIO). Utilizing the Kokkos performance portability framework and the ArborX bounding volume hierarchy (BVH) library, the entire Point cloud pipeline was moved onto GPU, which includes voxelization, deskewing, GICP and map handling.

Link to popular science summary to be added

15:15-16:00 in E:2116

• Presenters: Hugo Persson, Yazan Al-Aswad
• Title: High-performance, High-Throughput memory-efficient lookup tables for schemaless relational data
• Examiner: Per Andersson
• Supervisor: Jonas Skeppstedt (LTH)

Bulk importing data into Neo4j requires a temporary mapping from external identifiers to internal 64-bit identifiers to resolve relationships before the graph is materialized. At a billion-key scale, this lookup becomes a major bottleneck, yet no prior work has empirically compared modern retrieval structures in this setting.

This thesis benchmarks state-of-the-art lookup structures for the externalId-to-internalId mapping. Four architectures were implemented on top of BuRR and PtrHash, including hybrid designs that combine a coarse-grained well function with sharded per-well minimal perfect hash functions. They were evaluated against the BinarySearchLookup baseline at 10^9 and 10^10 keys on a Lenovo ThinkPad.

At 10^9 keys, Direct_PTR reduced end-to-end import time by 84% (6.1× speedup). At 10^10 keys, Hybrid_BuRR_PTR was 11% faster than Direct_PTR and the BinarySearchLookup is too slow to benchmark because of I/O page faults.

Here is a link to the popular science summary