BEGIN:VCALENDAR PRODID:-//eluceo/ical//2.0/EN VERSION:2.0 CALSCALE:GREGORIAN BEGIN:VEVENT UID:dbe9b3b33c455d739721bac1885e1a22 DTSTAMP:20260609T085749Z SUMMARY:CS MSc Thesis Presentation Day 4 June 2026 DESCRIPTION:Kontakt: birger.swahn@cs.lth.se\n\nThursday 4 June is a day for coordinated master thesis presentations in Computer Science at Lund Unive rsity\, Faculty of Engineering. Ten theses will be presented.You will find information about how to follow along under each presentation. The presen tations will take place in E:2116 and E:4130 (Lucas). A preliminary schedu le follows.Note to potential opponents: Register as an opponent to the pre sentation of your choice by sending an email to the examiner for that pres entation (firstname.lastname@cs.lth.se). Do not forget to specify the pres entation you register for! Note that the number of opponents may be limite d (often to two)\, so you might be forced to choose another presentation i f you register too late. Registrations are individual\, just as the opposi tions are! More instructions for opponents are found here on the LTH thesi s project page.Many more presentations will take place on 11 June\, see th is link\, and some on other dates see more on this link.09:15-10:00 in E:2 116Presenters: Hanxuan Lin\, Hanyu LinTitle: Multi-Stage Dashboard Generat ion over Graph Databases with LLMsExaminer: Emelie EngströmSupervisors: R ushali Gupta (LTH)\, Niels De Jong (Neo4j)\, Noah Mayerhofer (Neo4j)This t hesis investigates natural language–driven dashboard generation over gra ph databases\, with a focus on maintaining alignment between user intent a nd generated outputs. While prior work in natural language to visualizatio n has improved single-chart generation\, producing coherent multi-view das hboards remains challenging\, particularly for graph-structured data where relationships and query constraints are central. To address this\, the wo rk frames dashboard generation as a multi-stage workflow rather than a sin gle-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 implem ented for a Neo4j database and evaluated against a baseline generation met hod through a comparative dashboard study and an interactive workflow stud y. 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 summary09:00-10:00 in E:4 130 (Lucas)Presenters: Wenrui Xie\, Jiacheng ZhangTitle: Diffusion Policy for Force- and Vision-Guided Robotic Liquid Pouring TaskExaminer: Maj Sten markSupervisor: Volker Krueger (LTH)Liquid pouring is a high-precision\, c ontact-rich task that remains challenging for imitation learning in labora tory automation and pharmaceutical manufacturing. This thesis extends diff usion policy for dual-arm robotic pouring by integrating force feedback wi th visual observations. Diffusion policy is suitable for pouring because i t can model multi-modal action distributions and generate smooth temporal action chunks rather than isolated point-wise commands. To improve physica l awareness\, joint torque signals are added as proprioceptive inputs\, en abling the policy to respond to load and interaction changes during execut ion. For safety\, rice is used as a proxy medium instead of real liquid. T he policy is trained on 150 human demonstrations and evaluated over 60 tri als across three configurations: single-view baseline\, multi-view baselin e\, and force-integrated multi-view policy. Results show that the force-in tegrated policy achieves the highest overall success rate and better task progression. Generalization to real liquids remains future work.Link to po pular science summary to be added10:15-11:00 in E:2116Presenters: Julius A ndreasson\, Hussein TaherTitle: Impact of JVM Optimization on Energy Consu mptionExaminer: Niklas ForsSupervisor: Christoph Reichenbach (LTH)As the e nergy consumption of the IT industry grows\, so does the need for sustaina ble computing. This thesis investigates the impact of JVM tiered compilati on on energy consumption\, specifically comparing Tier 1 and Tier 4.The co llection of measurement data during workloads\, including power draw and J VM code execution\, was automated using a containerized system. Using this system\, we ran various benchmarks with different JVM configurations. Noi se was minimized through locked CPU frequencies\, disabled hyper-threading and kernel isolation. JVM code execution data was collected using DMCE in strumentation. We analyzed code heatmaps to correlate optimizations with e nergy-related impacts\, increasing our understanding.Our findings confirm that JVM configuration affects energy efficiency for workloads. We found t hat while power draw remains consistent between tiers\, Tier 4 significant ly 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 added11 :15-12:00 in E:2116Presenters: Ludvig Lindholm\, Danny TangTitle: Hierarch ical Self-Supervised Learning for Semantic Embeddings of Crash DumpsExamin er: Niklas ForsSupervisors: 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 a nd inefficient at Schneider Electric due to basic hashing methods. This th esis proposes a machine-learning embedding system that maps entire crash d umps into single semantic embeddings to automatically group them by underl ying defects.Since only a small fraction of the crash dumps are labeled\, we opted for primarily using self-supervised training. To process these la rge files\, we introduced a hierarchical encoder-based model that chunks d ata and combines it using learned\, adaptive weighting. We also developed a novel attention-based loss function and a multi-phase curriculum to prev ent overfitting.Our self-supervised model significantly outperforms the Mo dernBERT baseline\, nearly halving the false positive rate at a 95% retrie val threshold\, proving its viability for clustering. A supplementary supe rvised model showed further promise in real-world manual testing despite g round truth data scarcity.Link to popular science summary to be added11:15 -12:00 in E:4130 (Lucas)Presenters: Haobo Zu\, Yaliang CaiTitle: Graph-Aug mented RAG for Multimodal Document Question AnsweringExaminer: Xuan-Son Vu Supervisors: Pierre Nugues (LTH)\, Xuhao Zhang (Huawei Sweden R&\;D)Lar ge language models require grounded\, structured evidence to answer questi ons reliably over private\, multimodal enterprise documents. Existing grap h-based RAG systems address retrieval structure but rarely combine multimo dal document support\, genuine multi-hop graph traversal\, and adaptive re trieval control in a single deployable system. We present a graph-augmente d RAG system built on RAG-Anything and deployed locally to preserve data p rivacy. We improve knowledge graph coherence through entity resolution\, e nable multi-hop reasoning via Personalized PageRank-based traversal combin ed with hybrid vector retrieval and cross-encoder reranking\, and add an a gentic pipeline that routes queries adaptively\, detects retrieval insuffi ciency\, and verifies answer grounding. The system is deployed as an inter active web application. Experiments on DocBench Academic\, SurGE\, and thr ee multi-hop QA benchmarks show consistent improvements over the RAG-Anyth ing baseline.Here is a link to the popular science summary13:15-14:00 in E :2116 N.B. No more opponents for this presentationPresenters: Antonio Krso ski\, Emanuel SjövallTitle: Managing Failures in Stateful SystemsExaminer : Alma Orucevic-AlagicSupervisors: 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-base d environments\, where Operators automate the management of stateful appli cations but also introduce additional complexity and failure modes.This th esis investigates failures in Operator-managed stateful systems within Neo 4j’s cloud architecture. The initiating problem is that failures still o ccur despite existing testing and rollout strategies\, and recovery often requires manual\, non-standardized intervention. To address this\, we exam ine two research questions: what root causes lead to failures in managed s tateful systems\, and how failures caused by Kubernetes Operators can be m itigated.We identified 74 root causes and grouped them into 36 responsibil ity domains. Failures often stem from incorrect state evaluation and error -prone deployments. We developed Reconcile Dump to improve debugging and i dentified key operational pain points.These findings suggest that improvin g reliability requires incremental improvements in observability\, automat ion\, and deployment design.Here is a link to the popular science summary1 3:15-14:00 in E:4130 (Lucas)Presenters: Arvid Malm\, Wolmar Boris-MöllerT itle: An On-Premise Diffusion Deep Research Model &\; Evaluation Framew ork for Resource-Constrained EnvironmentsExaminer: Pierre NuguesSupervisor s: Patrik Edén (LTH)\, Thomas Drakengren (Qrendo AB)State-of-the-art deep research models for generating long-form\, citation-grounded reports typi cally rely on closed-source models and cloud infrastructure\, limiting the ir use in high-security environments. This thesis explores whether such sy stems can be deployed fully on-premise for requirements management using a design science research approach with two main artifacts.First\, we devel op an on-premise research agent\, Diffusion On-premise Research Agent (DOR A)\, 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 f actual validation across four text quality metrics and correctness.The res ults demonstrate that it is practically viable to run deep research archit ectures on-premise\, granted appropriate design choices are made and small er language models are used. The proposed model DORA outperforms all evalu ated 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 sum mary14:15-15:00 in E:2116Presenters: Emelie Tingberg\, Victor SannicoloTit le: Test-Driven Development as a method for AI-generated codeExaminer: Bj örn RegnellSupervisors: Per Runeson (LTH)\, Ivan Aladjoff (Decerno AB)Lar ge Language Models (LLMs) have created new possibilities for automating so ftware development\, but ensuring the quality of AI-generated code remains a challenge. This thesis investigates whether Test-Driven Development (TD D) can improve the quality of code generated by an LLM in a fully autonomo us context at Decerno AB.A TDD workflow was iteratively designed and compa red against a non-TDD workflow using static code analysis. TDD reduced cod e complexity and duplication but introduced more quality issues overall\, while non-TDD produced fewer but more severe issues. The analysis of gener ated tests in the TDD workflow revealed that the tests primarily focused o n content verification rather than behavioral testing\, contributing to a lower code coverage per test compared to non-TDD.The conclusion is that TD D quality depends on which development priorities are valued. This suggest s strict TDD may not be optimal in autonomous AI-driven contexts\, motivat ing more flexible workflow adaptations.Here is a link to the popular scien ce summary14:15-15:00 in E:4130 (Lucas) N.B. No more opponents for this pr esentationPresenter: Alve LindellTitle: GPU acceleration of LiDAR Inertial OdometryExaminer: Jonas SkeppstedtSupervisors: Michael Doggett (LTH)LiDAR Inertial Odometry (LIO) provides highly accurate state estimation for aut onomous robotic systems but suffers from significant computational bottlen ecks\, especially during the Generalized Iterated Closest Point (GICP) Poi nt cloud registration. On resource-constrained edge devices\, this high CP U demand can prevent real-time execution and constrain the available compu tational resources that could be required for other tasks. This thesis pre sents a GPU-accelerated implementation of Direct LiDAR-Inertial Odometry ( DLIO). Utilizing the Kokkos performance portability framework and the Arbo rX bounding volume hierarchy (BVH) library\, the entire Point cloud pipeli ne was moved onto GPU\, which includes voxelization\, deskewing\, GICP and map handling.Link to popular science summary to be added15:15-16:00 in E: 2116Presenters: Hugo Persson\, Yazan Al-AswadTitle: High-performance\, Hig h-Throughput memory-efficient lookup tables for schemaless relational data Examiner: Per AnderssonSupervisor: Jonas Skeppstedt (LTH)Bulk importing da ta into Neo4j requires a temporary mapping from external identifiers to in ternal 64-bit identifiers to resolve relationships before the graph is mat erialized. 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 implemen ted on top of BuRR and PtrHash\, including hybrid designs that combine a c oarse-grained well function with sharded per-well minimal perfect hash fun ctions. 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 en d-to-end import time by 84% (6.1× speedup). At 10^10 keys\, Hybrid_BuRR_P TR was 11% faster than Direct_PTR and the BinarySearchLookup is too slow t o benchmark because of I/O page faults.Here is a link to the popular scien ce summary \;\n\nMer information om händelsen: https://www.cs.lth.se/ evenemang/cs-msc-thesis-presentation-day-4-june-2026 DTSTART;TZID=GMT:20260604T070000 DTEND;TZID=GMT:20260604T140000 LOCATION:E:2116 & E:4130 (Lucas) END:VEVENT END:VCALENDAR