Open-domain question answering (ODQA) often requires integrating evidence from multiple sources and reasoning across several steps. While recent work has made progress on retrieval and reasoning independently, their combined optimization remains challenging. Standard retrieval methods may fail to surface all relevant documents, while reasoning models can generate confident but incorrect answers when evidence is incomplete, noisy, or inconsistent. This limits both accuracy and the reliability of model predictions, particularly for multi-hop and compositional questions.

This thesis explores strategies to jointly enhance retrieval and reasoning for ODQA. On the retrieval side, we introduce a dynamic answer frontier mechanism that prioritizes candidate documents based on semantic consistency across multiple generated answers, guiding iterative document expansion over a retrieval graph. This consistency-driven approach improves recall by promoting documents aligned with the most reliable reasoning traces. On the reasoning side, we apply test-time scaling (TTS), generating multiple candidate answers per question and training a verifier model to select the most trustworthy one. The verifier evaluates both semantic correctness and grounding in retrieved evidence, mitigating the effects of misleading or irrelevant documents.

We evaluate the proposed approach on two challenging ODQA benchmarks, MuSiQue and 2WikiMultiHopQA, which require complex multi-hop reasoning and resist shortcut solutions. Experimental results show that our method improves evidence recall and downstream answer accuracy over strong baselines, including standard retrieval pipelines and semantic uncertainty-based re-ranking methods. Qualitative analysis reveals better handling of compositional queries, including temporal comparisons and multi-hop relational reasoning, along with improved resilience to noisy retrievals and reduced divergence from relevant evidence. The study also identifies remaining challenges, such as reliance on agreement as a proxy for correctness and the computational cost of TTS, pointing to future directions involving principled uncertainty measures, end-to-end feedback integration, and efficiency improvements for exploring larger reasoning spaces. Together, these findings underscore the value of integrating semantic consistency-driven retrieval with verifier-guided reasoning selection to advance robustness and trustworthiness in complex ODQA systems.

Combining data from Randomized Controlled Trials (RCTs) is a widely used method to estimate causal treatment effects. In order to combine data, the property of transportability, under which different covariate vectors exhibit similar treatment benefit, must hold between the RCTs. However, differences in study design, execution, and the underlying effect modifier distributions can violate transportability which could in turn lead to estimating incorrect causal treatment effect estimates. This thesis addresses the challenge of validating transportability between multiple RCTs and identifying subsets of RCTs between which transportability holds. Our contributions include studying a linear regression-based framework for testing transportability between multiple RCTs and a clustering-based approach for identifying transportable RCT subgroups. Through simulations and analysis of real-world RCTs concerning corticosteroid treatment for Community-acquired pneumonia (CAP), we evaluate the power, robustness, and limitations of our proposed framework.

Coastal zones are dynamic and vulnerable regions, demanding accurate, scalable monitoring tools to inform environmental management and hazard mitigation. While satellite imagery and CNN-based classifiers have improved automated mapping, their reliance on unstructured pixel data limits contextual understanding. This study presents the first fine-tuning of a multi-modal large language model (MLLM), Qwen2.5, on 12-channel satellite input for multilabel coastal classification, demonstrating how architectural adaptation enables integration of spectral, topographic, and derived features beyond RGB. We compare this approach to a ResNet-50 baseline and state-of-the-art prompting methods using GPT-4o and LLaMA-3.2. Our experiments on the CoastBench dataset reveal that MLLMs benefit substantially from few-shot prompting with diverse, balanced sampling and that fine-tuning Qwen2.5 with full 12-channel input outperforms its RGB-only variant. An ablation study quantifies the importance of elevation and water-sensitive indices, while a human benchmark exposes a performance ceiling near F1 ≈ 0.70 due to label ambiguity. Our findings suggest that while MLLMs can rival traditional models and offer interpretability benefits, future gains depend on dataset quality, input diversity, and prompting strategy design.

Effective LLM-based automated program repair (APR) methods can lead to massive cost reductions and have improved significantly in recent times. However, the validity of many APR evaluations as they are conducted at this point is at risk due to data leakage: Prior research has shown that LLMs can memorize solutions to problems if the evaluation benchmark overlaps with the training set, leading to overinflated results.

In this study, we examine the potential of using metamorphic transformations to mitigate the effects of data leakage. For this, we create a variant benchmark for two popular, well-established benchmarks Defects4J and GitBug-Java, and evaluate the APR performance of several LLMs on these benchmarks and their transformed counterparts. In addition, we investigate to what extent our results align with data leakage metrics from other studies.

Our results show that state-of-the-art LLMs for code repair exhibit significant performance degradation (Up to 4.1% for Claude-3.7-Sonnet) on a metamorphically transformed Defecsts4J benchmark. Moreover, we find a significant correlation between our results and the negative log-likelihood as a metric of data leakage. Our results demonstrate the potential of using metamorphic transformations to mitigate the overinflation of evaluation results due to data leakage. We recommend that researchers report results on both original and metamorphically transformed benchmarks in future evaluations.

Writing test cases is an important yet complex task. Search-Based Software Testing (SBST) is an automated test case generation technique that aims to help developers by creating high-coverage test cases. Despite its strengths, a major limitation of this technique is that it often struggles with generating test cases that contain complex method sequences, as they have no semantic understanding of which methods are related. The recent advancement of Large Language Models (LLMs) offers a potential solution due to their natural language capabilities and applicability to software engineering tasks. However, LLMs often end up with lower coverage than SBST when directly compared due to lacking the output diversity that is required in software testing. This opens an opportunity to combine the exploratory power of SBST, with the semantic understanding of LLMs to make the test case generation process more effective in terms of test coverage and test structure. This thesis investigates the combination of these methodologies with our hybrid approach, LLM-Seeded Evolutionary Testing (LSET), which uses LLMs to generate tests that contain complex method sequences, and introduces this structure into the SBST process by serving as the starting point (seeds) of the algorithm to be evolved further. We conducted an empirical evaluation on a benchmark consisting of 35 JavaScript classes and found a significant branch coverage increase on 19 and 14 classes when compared to SBST and LLM-only baselines. However, when taking the combination of final SBST and LLM test suites, this gap reduces to 1 out of 35 classes. Beyond coverage, we also found a positive effect on structure when compared to tests generated by SBST, as the structure provided by the LLM tests can be further evolved to reach deeper branches while maintaining readability.

The increased use of mobile phones in security related applications has increased the need to verify device integrity. Consumers use smartphones as a form of online identification. Mobile phones provide law enforcement a useful surface for criminal prosecution. Manufacturers constantly patch vulnerabilities to prevent data leaks. Finding exploitable vulnerabilities, however, is non-trivial due to device encryption. One vector of attack is compromising a device driver to access privileged kernel information.

Finding exploits is difficult, time-consuming, and frequently requires in-depth knowledge of the surface under attack. Furthermore, software developers and manufacturers are continuously patching vulnerabilities and upgrading the interface. This makes finding vulnerabilities prone to errors, even for experts.

This thesis focusses on automating the process of finding vulnerabilities in Android device drivers. Several tools exist that automate part of the process, such as Syzkaller and the Evasion kernel. However, each individual tool leaves gaps in their use that make them impractical for realistic situations. Syzkaller is able to fuzz the Linux kernel, but often lacks the necessary components for fuzzing device drivers. The Evasion framework can emulate Android device drivers, but fuzzing these drivers requires in-depth knowledge of their internals.

Therefore, this thesis presents FELIX: a novel toolchain that is able to instrument and fuzz Android device drivers in an emulated environment. First, FELIX instruments the device driver and kernel in order to emulate the drivers without meeting the hardware requirements. Second, FELIX analyses the device driver to create the interface for a fuzzer. Lastly, FELIX uses Syzkaller to test the driver for vulnerabilities or exploits.

FELIX successfully fuzzed five different Android device drivers. In doing so, FELIX was able to reproduce known vulnerabilities, and managed to reach code that was previously uncovered. This demonstrates the ability of FELIX to discover new vulnerabilities in the future.

Electropherograms (EPGs) are produced in forensic laboratories when DNA traces are collected and processed. EPGs produced under good conditions with a single DNA contributor are relatively easy to interpret. With more contributors, EPGs are difficult to interpret fully correctly due to overlapping peaks, variability across loci and runs, proliferation of stutters & other artefacts. These challenges are amplified in DNA mixtures with low-template or imbalanced contributors, where noise-signal ratio is high. Here, traditional rule-based and statistical approaches alone fail, and are dependent on human analysts for the final calls. Recent deep learning methods show promise but remain limited by scarce verified training data, ambiguous ground-truth labels, and poor generalization across kits and genotypes.

This thesis investigates how data augmentation and preprocessing can improve the robustness of deep learning models for allele detection from EPGs. First, a conditional generative adversarial network (cGAN) is employed to generate synthetic yet realistic electropherograms, extending training diversity beyond the limited set of verified contributors. Second, preprocessing techniques are introduced standardize EPG data. The key pre-processing technique, Scan Point Standardization (SPS), aligns signals across runs using the internal lane standard, complemented by dye-wise fluorescence (RFU) standardization and dynamic-range compression (DRC) to mitigate intensity imbalance and baseline noise. Third, a genotype-aware data-splitting protocol is developed using spectral clustering to prevent genotype leakage between train and test sets, ensuring fair generalization assessment. Lastly, the performance of models developed in this research are applied to NFI's R&D data, which uses a different EPG kit in order to investigate the cross-kit compatibility of DL models.

Experimental results show that introducing synthetic data alone did not improve performance on our dataset. On the other hand, of the preprocessing techniques, none of standardization, DRC SPS improved performance for pixel metrics. Genotype-aware splitting reveals that conventional random file-level splits significantly overestimate model performance by allowing for train-test leakage. Once SPS and genotype-aware leakage-free splitting were incorporated, synthetic data then results in significant performance improvements. Collectively, these findings demonstrate that preprocessing, realistic synthetic augmentation, and leakage-free evaluation are essential for achieving robust, generalizable allele calling in forensic DNA profiling. Models trained on purely synthetic data as well synthetic & real data using the GlobalFiler kit were then trained and tested on NFI's R&D data using the PPF6C kit, showing that transfer learning is possible between different PCR kits.

Computers have become an essential part of modern life. They are used in a wide range of applications, from smartphones and laptops to data centers and supercomputers. However, the increasing usage of computers has led to a rise in energy consumption, which has significant environmental and economic consequences.

The hardware of computing systems has become more energy-efficient over the years. However, software remains a largely untapped area for energy optimization. The key to reducing energy consumption in computing systems lies in understanding the impact of software on the overall energy usage of the system. So, the need for tools and methods to measure software energy consumption is crucial. Those tools and methods should be flexible enough to be used in different computing environments.

This thesis reviews the current state-of-the-art tools for monitoring software energy consumption. After a list of their capacities, we identify the more promising tools and methods and compare their performance during experiments using different types of workloads. The thesis also presents a new tool called Geryon, which brings together the best techniques found in the tools reviewed. Geryon is a flexible and easy-to-use tool that provides multiple estimations at once. Those estimations are computed via simple power models and regression models trained on-the-fly. It also provides a way to plug in an external power model, allowing for more complex power models to be used. The tool is evaluated on the same workloads as the other tools to provide a fair comparison.

Procedural Content Generation (PCG) is a method to automatically generate content with little to no human assistance required. It emerges as a promising tool to generate educational content tailored to individual learning needs, a fundamental aspect of effective teaching.
This review aims to provide an overview of existing research on the educational applications of PCG across disciplines. The key objectives of this paper include identifying the range of domains that have profited from the educational use of PCG, the challenges faced by specific disciplines as well as discerning the similarities and differences in their approaches. By addressing these objectives, this literature review seeks to identify areas for further research and provide insight into understanding the obstacles preventing the adoption of PCG in certain educational domains. By carefully surveying a sample of research conducted on the educational use of PCG in a broad range of disciplines, we found that computer science and mathematics have extensively leveraged PCG to enhance learning experiences, while natural sciences and social sciences could benefit from further research. To the best of our knowledge, fields like geography, physics and biology as well as arts, business and economics remain largely unexplored. Furthermore, challenges such as adjusting content difficulty, generating coherent exercises, and educational accuracy issues have been identified as significant barriers in various fields.

Procedural Content Generation (PCG) is a powerful content generation technique that can be used to automatically generate content (an example would be exercises for a quiz or a game). As it stands, PCG is able to create content with zero human interaction which makes it a technique worth exploring for educational purposes. In the context of PCG, this research focuses on how PCG is used to orchestrate the simultaneous generation of content of various types and identify what could be done more, possibly towards educational use. The method chosen for this research is literature review. As such, this paper provides a summary of "orchestrative" PCG approaches found in recent literature and evaluates said approaches through the lens of "PCG for education". Findings are presented in an exhaustive manner with each piece of literature occupying a subsection. A discussion section that explores how these approaches could benefit education and/or what could be done to facilitate that is also present.

The use of deep learning models has advanced in gaze-tracking systems, but it has also introduced new vulnerabilities to backdoor attacks, such as BadNets. This attack allows models to behave normally on regular inputs. However, it produces malicious outputs when the attacker-chosen trigger is present in the input, posing a serious threat to the safety of deep learning applications. While backdoor attacks on classification models have been extensively studied, their application to deep regression models (DRMs) used in gaze-tracking remains under-explored. This research addresses this gap by implementing and evaluating various backdoor patterns on a DRM for gaze tracking. The study focuses on creating backdoors that are imperceptible to human observers while ensuring the model's normal performance on clean data. Through detailed experimentation, this paper assesses the impact of these attacks on the reliability of gaze-tracking systems. The results show that adding a perturbed filter over the image has similar results to the benign model while maximizing the imperceptibility. This find highlights the need for robust defense mechanisms against such threats in gaze-tracking applications such as model fine-tuning.

This research investigates backdoor attacks on deep regression models, focusing on the gaze estimation task. Backdoor triggers can be used to poison a model during training phase to have a hidden misbehaving functionality. For gaze estimation, a backdoored model will return an attacker-chosen target gaze direction, normally incorrect, regardless of image content, when presented with an image containing a trigger. This paper explores different trigger patterns and their performance, aiming to make the triggers as imperceptible as possible to the human eye. Furthermore, the research explores a method to make the corruption of the training set as stealthy as possible while achieving a good attack performance. In the end, the findings showed that backdoor attacks on deep regression models can be made imperceptible and highly performant using complex trigger patterns. While stealthy corruption was also possible, achieving an efficient model would require a larger dataset.

With the rise of deep learning and the widespread use of deep neural networks, backdoor attacks have become a significant security threat, drawing considerable research interest. One such attack is the SIG backdoor attack, which introduces signals to the images. We look into three types of SIG backdoor attacks - ramp, triangle, and sinusoidal signals. Most of the works in the field of AI security, however, have focused on deep classification tasks, leaving deep regression tasks unexplored. In this study, we adapt the SIG backdoor attack for use in a deep regression model (DRM) used to estimate head pose. Our objective is to create a backdoor attack that remains imperceptible to the human eye while being detectable by the DRM. To evaluate the effectiveness of our attack, we employ two approaches: average angular error and accuracy in a discretized continuous space. Additionally, we adapt fine-tuning as a countermeasure against the backdoor attack. By implementing this strategy, we aim to reduce the risk of backdoor attacks and improve the robustness of deep regression models in head pose estimation.

Deep Regression Models (DRMs) are a subset of deep learning models that output continuous values. Due to their performance, DRMs are widely used as critical components in various systems. As training a DRM is resource-intensive, many rely on pre-trained third-party models, which can leave a worrying amount of systems vulenerable to backdoor attacks. A backdoored model is an otherwise legitimate model that maliciously changes its behaviour whenever a predetermined backdoor trigger is present. While numerous works on backdoor attacks on deep learning models focus on classification problems, very little work has focused on DRMs. We formulate and evaluate a backdoor attack on a DRM using WaNet, a method that relies on warping-based triggers that are difficult to detect by both human and machine defence methods. We successfully train a backdoored (poisoned) DRM with the backdoor working for both grayscale and coloured inputs. Further experiments show that the malicious backdoor behaviour can be subdued by fine-tuning the poisoned model.

Previous research has explored the detection of adversarial examples with dimensional reduction and Out-of-Distribution (OOD) recognition. However, these approaches are not effective against white-box adversarial attacks. Moreover, recent OOD methods that utilize hidden units hinder the scalability of the target model.

For that reason, various explanations of adversarial examples are studied to get a better understanding about its properties and anomalies. Furthermore, we discuss the added value of using natural scene statistics and utility functions to improve the relevance of the features for detection. By utilizing the anomalies we identified for adversarial examples in an ensemble, this thesis is the first to propose a robust solution for adaptive and white-box attacks.

Particularly, we address these challenges with MeetSafe. A Gaussian Mixture Model that leverages principal component analysis, feature squeezing, and density estimation to detect adaptive white-box adversaries. Furthermore, our enhanced Local Reachability Density (LRD) algorithm further improves the efficiency of state-of-the-art OOD methods. In particular, the proposed LRD enhances scalability by feature bagging hidden units with large absolute Z-scores. We then show that predictors, including LRD, are far more effective in ensembles like MeetSafe which supports prior conjectures that a range of different heuristics may further constrain adversaries when combined.

Extensive experiments on 14 models show that MeetSafe detects adaptive perturbations with an accuracy of 62% on STL-10, 75% on CIFAR-10, and 99% on MNIST using either adversarial training or Reverse Cross Entropy (RCE), achieving an improvement of at least 8.1% for each evaluated method by averaging across the three datasets.

Emotion recognition is a challenging problem in the field of computer vision. The automatic classification of emotions using facial expressions is a promising approach to understand human behavior in various applications such as marketing, health, and education. How- ever, recognizing some emotions, such as anger, jealousy, contempt, and disgust, is more challenging than others due to their subtlety and rarity in the training data. In this paper, we try to investigate if using (self)pseudo-labelled data to train an Expression Manipulator [? ] generator to generate a training set for training a classifier is a better alternative to directly using an equal amount of (self)pseudo-labelled data for training the classifier [? ]. Specifically, we focus on augmenting the Action Units (AUs) of facial expressions, which are the basic units of facial movement that correspond to specific emotions

Chess recognition refers to the task of identifying the chess pieces configuration from a chessboard image. Contrary to the predominant approach that aims to solve this task through the pipeline of chessboard detection, square localization, and piece classification, we rely on the power of deep learning models and introduce two novel methodologies to circumvent this pipeline and directly predict the chessboard configuration from the entire image. In doing so, we avoid the inherent error accumulation of the sequential approaches and the need for intermediate annotations.

Furthermore, we introduce a new dataset, Chess Recognition Dataset (ChessReD), specifically designed for chess recognition that consists of 10,800 images and their corresponding annotations. In contrast to existing synthetic datasets with limited angles, this dataset comprises a diverse collection of real images of chess formations captured from various angles using smartphone cameras; a sensor choice made to ensure real-world applicability. We use this dataset to both train our model and evaluate and compare its performance to that of the current state-of-the-art. Our approach in chess recognition on this new benchmark dataset outperforms related approaches, achieving a board recognition accuracy of 15.26% (≈7x better than the current state-of-the-art).

Kotlin is a programming language best known for its interoperability with Java, as well as the measurable improvements it offers over it. Since it became Android’s go-to language in 2019, the popularity and impact of Kotlin have risen greatly. Amidst this surge in popularity, the Kotlin developer team is working on a new version of the compiler that introduces sweeping changes to the ecosystem. Traditional compiler testing is a manual and laborious task that requires extensive developer effort and expertise. In an attempt to mitigate this, researchers have invested great resources in developing and perfecting automated compiler testing tools over the last decades. These approaches generate new pieces of code to test the behavior of compilers, which is assessed through differential testing. However, the usage of heuristics as guidance for the generative process is not well understood, and no approach that generates Kotlin code from scratch currently exists. In this thesis, we propose a novel method of enriching standard grammar specifications with language-targeted semantic context that is integrated in the sampling process. We structure generated code hierarchically and use it as the base of an evolutionary computation framework. Within this framework, we introduce two classes of algorithms that are novel to the field of compiler fuzzing, based on syntactic diversity and semantic proximity, respectively. We carry out an empirical analysis spanning 200K generated Kotlin files, which we analyzed through different Kotlin compiler versions. Our results uncovered five previously unreported categories of bugs, which we reported to the Kotlin compiler developer team. The developers verified and replicated our instances on the current re- lease of the Kotlin compiler, and have assigned target release dates for fixes within the current major version of the compiler. The study also provides new insight into the effects of heuristic-specific hyperparameters such as expression simplicity, dissimilarity measurements, and target selection.

With the ever-increasing digitalisation of society and the explosion of internet-enabled devices with the Internet of Things (IoT), keeping services and devices secure is becoming more important. Logs play a critical role in sustaining system reliability. Manual analysis of logs has become increasingly difficult, accelerating the development of automated methods for log-anomaly detection. Despite significant progress in automating log analysis, current state-of-the-art methods face challenges dealing with unstable log data, which means that the content of log messages evolves over time.

We show that LogBERT, a state-of-the-art technique based on Bidirectional Encoder Representations from Transformers (BERT), cannot deal with unstable log data. On the three most prevalent publicly available log datasets, Mathew's Correlation Coefficient (MCC) score (which measures the correlation between a model's output and the correct labels) of LogBERT dropped by 90% after increasing log data instability from 1% to over 80% normal sequences containing logkeys in the test set. Log data instability was increased by only reassigning samples between the train and test set. Furthermore, we show that the high performance of LogBERT reported in the original paper was achieved because the model relied on a simple heuristic that only worked under specific conditions.
To address this issue, we propose a novel sequence anomaly detection technique based on BERT: Vocabulary-Free BERT (VoBERT). VoBERT uses a novel pre-training task we designed specifically for anomaly detection: Vocabulary-Free Masked Language Modeling (VF-MLM). We adapted traditional MLM and removed the fixed vocabulary constraint, which allows VF-MLM to classify out-of-vocabulary logkeys correctly.
We highlight that VoBERT is more stable than LogBERT and outperforms the latter in certain situations where log data is very unstable. For the public datasets, the MCC score of the specific train-test split used in the LogBERT paper dropped by 90% after reassigning the train-test split, increasing log data instability. In addition to sequence-level anomaly predictions, we evaluated all approaches on element level, providing a more granular performance assessment.
To assess the generalisation of the experimental results to real-world scenarios, we conducted a case study evaluating the anomaly detection models on real-world security event data collected at a large bank (50,000+ employees). We found that the simple heuristic did not work for this real-world data, having a negative correlation with the correct results. VoBERT showed performance on par with LogBERT on this real-world security event dataset.
We urge future researchers to evaluate their methods on real-world data, as we showed that the commonly used public datasets do not represent real-world scenarios. Furthermore, it is important to assess how difficult it is to detect anomalies in datasets used for evaluation. When a simple heuristic can perform well, such datasets might not be well suited to evaluate a complex anomaly detection model.
This thesis is a proof of concept for the novel pre-training task VF-MLM and paves the way for future work to refine this technique further, as well as to develop additional robust and adaptable solutions for log and security event anomaly detection.

On the intuitive level, software testing is important because it assures the quality of the software used by humans. However, ensuring this quality is not an easy task because as the complexity of the software increases, so do the efforts to test it. Search-based software testing is an active research field that develops and explores tools for automatic test case generation. Their work involves using meta-heuristic search optimisation approaches such as evolutionary algorithms
from the evolutionary computation community and applying them to test case generation. The crossover between the evolutionary computation domain and the test case generation one produced DynaMOSA, a state-of-the-art evolutionary algorithm for generating test cases. In an attempt to
produce another well-performing algorithm, this paper performs another crossover between the two communities and creates DynaMOSARVEA - a product of a Reference Vector Guided Evolutionary Algorithm (RVEA) and DynaMOSA.
The conducted empirical study showed that although DynaMOSARVEA does not outperform DynaMOSA, it did outperform RVEA, thus demonstrating the value brought by domain-specific knowledge.