Retrieval-Augmented Generation (RAG) extends Large Language Models with external knowledge retrieval, yet the scaling behavior of inference-time compute in multi-stage RAG pipelines remains poorly characterized. This thesis provides a systematic empirical investigation of test-time compute scaling along three axes: retrieval depth (expanding the candidate pool size c presented to a neural reranker), query breadth (generating n LLM-produced query reformulations and fusing their results), and adaptive budget allocation (dynamically distributing cross-encoder compute via ORE and ReformIR). Experiments span three benchmark datasets (MS-MARCO, TriviaQA/DPR-w100, 2WikiMultiHopQA), two first-stage retrievers (BM25, TCT-ColBERT), six fusion algorithms, and two pipeline orderings (merge-then-rerank vs. rerank-then-merge).

Our findings establish six empirical principles for test-time compute allocation in compound retrieval systems. First, depth scaling yields monotonic gains only when the reranker is architecturally robust to noise; the generative reranker MonoT5 degrades at large c while the distilled cross-encoder MiniLM-L6 scales reliably. Second, query breadth scaling requires a continuous embedding space, dense retrievers enable substantially larger fusion gains than sparse retrievers because semantically distinct rewrites activate genuinely different document regions. Third, the optimal pipeline ordering depends on the reranker's domain calibration: rerank-then-merge excels in-domain, while voting-based fusion dominates out-of-domain. Fourth, adaptive and fusion-based scaling are complementary rather than competing. Fifth, a persistent 15-17% relative Recall@50 Oracle gap reveals that practical fusion algorithms, not candidate generation, constitute the performance bottleneck. Sixth, retrieval gains transfer only partially to downstream generation quality, with volatile LLM-as-Judge trajectories cautioning against assuming proportional end-to-end improvements.

These results demonstrate that, unlike the monotonic scaling laws governing language model pre-training, retrieval scaling is non-monotonic and conditional, effective test-time compute allocation in Compound AI Systems requires domain-aware orchestration across the full retriever-reranker-fusion-reader pipeline. ... Read more

Machine-learning-driven management of next-generation (6G) networks depends on measurements that are routinely corrupted by sensor noise, hardware imperfections, bursty interference, and malicious activity, so outlier handling is widely assumed to be a prerequisite for reliable downstream models. Whether, and which, cleaning methods actually help, and whether this differs across data modalities, remains unclear. Using two real, labelled datasets with no synthetic contamination, attack traffic from a functional 5G testbed (attack classification) and an operational web-latency KPI series (short-term forecasting), we systematically compare six outlier-handling methods, namely interpretable statistical filters (robust Z-score replacement, IQR clipping, and Savitzky–Golay smoothing) and unsupervised detectors (Isolation Forest, Local Outlier Factor, and PCA reconstruction), against a no-cleaning baseline. Hyperparameters are tuned without access to held-out labels; each method is evaluated under both a robust (Random Forest) and a noise-sensitive (k-NN) downstream model, with paired significance tests and false-discovery-rate (FDR) correction, a detection diagnostic, and runtime. The result is largely negative: after FDR correction, no method significantly improves downstream performance on either modality. Savitzky–Golay smoothing gives the only suggestive forecasting gain (≈17% lower error under Random Forest) but does not survive correction; deletion- and clipping-based methods are neutral-to-harmful (IQR significantly degrades classification); and the unsupervised detectors rank real attacks barely above chance (ROC-AUC 0.54–0.60), even though a supervised model separates the same classes at 0.86, statistical outlier detection is a poor proxy for the anomalies of interest. As the slowest detectors are also the most harmful and exceed the near-real-time control budget, we conclude that a generic outlier-handling stage offers no reliable benefit for these tasks: its value must be demonstrated rather than assumed, with lightweight smoothing the only candidate worth trying on noisy sequential signals. ... Read more

Sixth-Generation (6G) telecommunications rely on high-frequency millimeter-wave (mmWave) bands for massive data rates, but their physical fragility makes them highly susceptible to line-of-sight blockages. These blockages cause contiguous telemetry outages, creating a single point of failure for edge routing and orchestration protocols demanding continuous system data. To address this, we introduce an evaluation pipeline benchmarking five time-series imputation architectures, from statistical baselines (Nearest Neighbor, Kalman Filter) to complex deep learning models (BRITS, CSDI, TimesNet). Utilizing an open-source microservice dataset, the pipeline dynamically injects simulated blockages across a 24-scenario grid, escalating from minor drops to 60-second outages. Performance is evaluated across an accuracy-latency Pareto frontier. Results demonstrate that the recurrent architecture, BRITS, achieves the highest overall reconstruction fidelity. However, Nearest Neighbor emerges as the optimal low-latency baseline, maintaining competitive accuracy while consistently executing in under 250 milliseconds. Finally, contextualizing these findings reveals a critical limitation: the architectures achieving peak accuracy inherently rely on offline, bidirectional processing to reconcile telemetry gaps. This highlights a significant research opportunity, emphasizing the need to evaluate deep learning models in strictly online, forward-only forecasting configurations to meet the split-second streaming realities of live 6G edge deployment. ... Read more

Sixth-generation (6G) networks collect positioning data that must be transformed into a suitable representation before machine-learning models can use it effectively. The choice of this encoding is rarely treated as an experimental variable, yet it strongly shapes what information reaches the downstream model. This paper evaluates how tabular and time-series encoding techniques affect beam prediction performance and feature representation quality in nine scenarios from the DeepSense 6G dataset. Beam-prediction performance is measured using two downstream classifiers in a fixed multi-seed evaluation pipeline, while representation quality is assessed through invariance under positional noise. Encodings that represent the user equipment relative to the base station and include temporal context achieve the best performance. However, the representation analysis reveals that these geometry-aware encodings are less stable under positional noise. The findings suggest that, when position estimates are accurate, position and trajectory data should be encoded using base-station-relative distance, bearing and recent geometric change, whereas noisier settings may require additional preprocessing to preserve robustness. ... Read more

Sixth-generation (6G) wireless systems depend on data-hungry machine-learning pipelines, yet datasets collected from heterogeneous sources frequently contain missing values that bias models and degrade simulation reliability. Tabular imputation has been studied extensively— from statistical baselines (mean, kNN) through model-based methods (MICE, SoftImpute) to recent deep approaches (HyperImpute, GRAPE, DiffPuter)—but no prior work systematically compares this range on 6G data under realistic missingness. We benchmark seven methods on DeepSense 6G datasets across four mechanisms and three missingness rates, evaluating reconstruction accuracy, statistical fidelity, and downstream beam-prediction performance. Our benchmarks show that no single imputation method consistently dominates; performance depends on the missingness mechanism. Under cell-wise missingness, deep methods such as HyperImpute achieve the highest reconstruction fidelity, though downstream beam prediction remains robust to these localised corruptions. In contrast, row-wise missingness degrades all learned and deep approaches by breaking cross-feature dependencies. Here, kNN is the only method that consistently preserves the downstream label signal. Overall, our results provide guidance for 6G pipeline defaults and highlight the limitations of applying purely tabular imputation to temporal wireless data. ... Read more

Data-driven methods are expected to play an important role in future sixth-generation (6G) wireless systems, where network data can support performance prediction, simulation, and network optimization. However, collecting large and representative network-performance datasets can be difficult, which motivates the use of data augmentation. This study evaluates how different tabular and time-series augmentation techniques compare when addressing data scarcity in datasets relevant to future 6G systems. Two regression tasks are studied: a tabular AMF performance task using XGBoost and a time-series Python web-server performance task using an LSTM. Four tabular augmentation methods are evaluated: Gaussian Noise, SMOGN, CTGAN, and TVAE. Four time-series augmentation methods are evaluated: Jittering, Time Warping, TS-Mixup, and Frequency-domain augmentation. The methods are compared using downstream regression performance, statistical realism metrics, and diagnostic analysis of augmented data and test-set residuals. The results show that augmentation does not consistently improve regression performance. In the tabular task, all augmentation methods reduced performance compared with the XGBoost baseline. In the time-series task, Frequency-domain augmentation was the only method that improved the LSTM baseline, substantially reducing RMSE and MAE, although the final test-set $R^2$ remained negative. The diagnostics suggest that useful augmentation depends not only on preserving marginal distributions or value ranges, but also on preserving task-relevant feature-target relationships and temporal structure. Overall, the findings show that augmentation effectiveness is method- and data-type dependent, and that predictive performance should be evaluated together with statistical fidelity diagnostics. ... Read more

Rendering volumes using path tracing can produce stunning images, but the process is notoriously expensive. For some volumes such as clouds, there are cases where the volume has no interaction with other objects in the scene. For these cases, combined with an isotropic phase function, unidirectional path tracing from the camera computes many similar paths regardless of camera position. Computing these light paths beforehand to determine the full radiance transport through the volume speeds up the rendering stage drastically, with the result still being physically accurate.

In this thesis, we present a two-stage method designed for the aforementioned cases. The first stage is a precomputation stage, in which light paths are traced through the volume and stored in a space-efficient manner using a graph structure. Through this graph, radiance transport is then computed. In the rendering stage, the graph structure with radiance values can then be used to render the volume. The results show that our method can produce renders with negligible bias. They also show that the method is heavily constrained by space requirements as volume size increases, which results in an increase in bias for larger volumes. The main advantage of the method is its efficiency in rendering multiple images of the same object. ... Read more

To bridge the offline-online evaluation gap, we in- vestigate the use of large language models (LLMs) for scalable user simulation. While LLMs offer a promising approach, the reliability of such simu- lations is highly sensitive to prompt design. This study makes two main contributions: (1) we intro- duce a novel three-dimensional reliability frame- work for evaluating LLM-based user simulations, capturing stability, discrimination, and plausibil- ity; (2) we systematically compare three prompt- ing strategies—One-Shot, Multi-Shot, and Offline- Evaluation-informed—across 52 retrieval systems. Results show that Multi-Shot prompting achieves the best balance across all reliability dimensions, producing simulations that are both consistent and sensitive to system quality. In contrast, Offline- Evaluation-informed prompts align closely with of- fline metrics but fail to differentiate system perfor- mance. These findings provide practical guidance for designing reliable LLM-based user simulators, reducing reliance on costly online experiments. ... Read more

Offline evaluation of Information Retrieval systems with its benefits of being overall cheaper than online evaluation still suffers from data sparsity. It has been established that LLMs can be used for labeling relevance much cheaper and faster than human annotators, while providing similar results. However it has not been measured how those labels affect the alignment between offline and online evaluation. In this work we present a labeling pipeline for a subset of queries from MS Marco. We then compare the labels against the baseline of Rank-DistiLLM dataset. We finalize by measuring alignment against external online metrics. In the results we observe that the metrics show strong correlation to both TREC 2019 and Rank-DistiLLM baselines. Measured alignment to Online metrics is moderate; However negative correlation present requires additional analysis.

https://huggingface.co/datasets/SemiGlassFace/CSE3000 ... Read more

Transformer-based architectures have significantly advanced the field of Information Retrieval (IR) by enabling semantic understanding that surpasses traditional term-frequency models. Hybrid approaches, which combine efficient sparse retrievers like BM25 with effective dense retrieval methods are becoming more popular for balancing performance and efficiency. However, the application of advanced hybrid systems often requires dedicated infrastructure and technical expertise, as existing toolkits are primarily research-oriented and not designed for simple integration.
To bridge this gap, this work introduces an open-source Elasticsearch plugin that implements the interpolation-based re-ranking framework from Leonhardt et al. This plugin simplifies the adoption of hybrid re-ranking by making the technique easily accessible within existing pipelines.
Furthermore, this work explores learning-based optimization methods to optimize the interpolation constant α. The research investigates two main strategies: (1) direct, gradient-based optimization to determine an optimized global value for α , and (2) the development of lightweight adaptor models that dynamically predict context-aware α values for each query or query-document pair. These adaptors are based on Feedforward Neural Networks and Neural Tensor Network architectures.
Empirical results suggest that these learning-based methods, particularly the dynamic adaptors, can outperform traditional grid search methods for tuning α, due to their dynamic adjustment in a query/document basis. The benefits were especially strong in out-of-domain scenarios, where the adaptors showed improved performance for encoders that were not pre-trained on the target domain, without requiring fine-tuning of the full backbone architecture.
This work delivers a practical, easy-to-use tool for hybrid re-ranking in Elasticsearch, a novel methodology for optimizing its core re-ranking parameter, and proposes adaptor models that can deliver better performance than a fixed interpolation value in hybrid re-ranking. ... Read more

Computer networks are an integral part of our society and functioning without them is difficult, as computers rely on this connection for their data or shared computing power. While this connectivity is often beneficial, it has downsides as well. A malicious actor can try to break into a network remotely, which organisations try to prevent by monitoring their networks in order to detect such an attacker. Monitoring can be done, for example, by searching for Indicators of Compromise (IoC) within the network traffic. These IoC can take the form of a single attribute, such as an IP address, or a combination of multiple attributes, such as an IP address, a protocol and a domain name. If during this search IoC match with the network data, a malicious actor might be present in the network. An in-house solution is not feasible for all organisations as this would result in more financial overhead, thus a managed Security Operations Center (mSOC) can be contracted. Often, such an mSOC has access to all network data in order to match their IoC. However, this may be undesirable for organisations that want to keep their sensitive network data as private as possible. Therefore, sharing relevant data only when a match has been found is preferable. Additionally, an mSOC has reasons to want to keep their IoC private as well, as they invest resources into gathering these IoC and sharing them would pose a risk to their business model. In this work we aim to match IoC, consisting of a variable number of attributes, with network data in order to retrieve data associated with matches while preserving the confidentiality of the unmatched data of both the mSOC and the organisation. While there exist privacy-preserving solutions that can aid in parts of this problem, no solution yet exists, to the best of our knowledge, that efficiently solves the problem entirely with our constraints on confidentiality.

To this end, we propose two privacy-preserving protocols that enable exact matching of variable multi-attribute IoC and network data. For both protocols, we analyse the theoretical complexity and test proofs of concept in order to highlight their strengths and weaknesses. ... Read more

This research addresses the challenge of deploying real-time drinking gesture detection in messy, "in-the-wild" environments. We propose and evaluate two computationally inexpensive systems, one using a Random Forest classifier, another using a 1-Dimensional Convolutional Neural Net (1D-CNN) classifier. Both are trained on 2D skeleton data, or features derived from that solely that data. Tested on the Conflab social interaction dataset, our method is designed to handle sparse labels and significant data occlusion. This study reports on the performance of this light-weight, video-based approach, providing a benchmark for applicability in real-world health and human-computer-interaction applications where privacy and computational efficiency are important factors. Although we were unable to create a robust and reliable classifier (f1 of 0.07 and 0.03 respectively), this work shows that there is potential for future work to succeed (roc-auc’s of 0.63 and 0.55 respectively) and provides critical insights into pitfalls to avoid when designing similar systems. ... Read more

Rankings are simply orderings given to a set of elements; They are a widely used mathematical object in information retrieval. This creates the need for some means of comparing them. Rank Similarity Measures are used exactly for this. They constitute a large research area where many different such measures are defined. A ranking may possibly contain ties. This in turn raises the question of what these ties represent and how to treat them in the calculation of a measure. The treatment of ties in current theory is approached with the a and b variants of the measures. Both a and b stem from a statistical approach to ties; they consider tied elements to represent uncertainty about their real order in the ranking. There is, however, a different interpretation of what ties could represent, namely that the tied elements really occur at the same place in the ranking, that is, there is no intrinsic order in which they should appear. This has been considered in one of the nonconjoint measures and has been coined the w-variant after Weber et al. In this work, we consider the problem of defining this very variant for a family of three commonly used ranking similarity measures, these being τ defined by Kendall, τ_AP defined by Yilmaz et al., and τ_h defined by Vigna. We approach this problem by establishing what the variant should represent and defining a set of axioms that any definition of w has to follow. Thereafter, we show that there is only one definition which can possibly satisfy these, with a small exception. We show that this definition coincides with the distance considered by Kemeny in 1959. We use this to create a definition of the w-variant for all three of the measures. Likewise, we investigate the behaviour of this new variant in relation to the existing a and b variants. Moreover, we identify the shortcomings of our definition and evaluate it on real world data. Finally, we lay the groundwork for rigorously proving parts of our definition and other measures which may consider ties to represent occurrence at the same rank. ... Read more

Adaptive retrieval is a technique to overcome the recall limitations of two-stage retrieval pipelines. Adaptive retrieval focuses mainly on effectiveness, but shows potential to improve efficiency. This research focuses on the trade-off between effectiveness and efficiency in adaptive retrieval. We explore the behaviour of neighbourhoods formed by the corpus graph, find that the effectiveness of adaptive retrieval varies across queries, and identify the most effective way to leverage the corpus graph. We investigate the impact of two different scoring mechanisms on the efficiency and effectiveness of an adaptive retrieval pipeline. We observe that score interpolation tends to improve adaptive retrieval's effectiveness, and incorporating an additional cross-encoder stage can lead to further gains. We compare our proposed solutions against several baselines to examine the trade-offs between effectiveness and efficiency. In our setting, we find that adaptive retrieval can improve efficiency at the cost of effectiveness. For our experiments, we introduce a visualisation tool for graph exploration and an adaptive retrieval component for efficient retrieval pipelines. ... Read more

A central problem in information retrieval (IR) is passage ranking, where the task is to retrieve passages from a corpus and order them in decreasing relevance to an arbitrary search query.
Traditional lexical retrieval methods are susceptible to the vocabulary mismatch problem, where relevant passages are overlooked if they do not contain the exact query terms (e.g., synonyms), despite being semantically relevant.
A recent trend in IR is to address this issue by utilizing neural network models (dense rankers) which embed text sequences into dense vector representations that effectively capture their semantics through complex attention mechanisms.
For efficiency, dense rankers are often employed in a retrieve-and-re-rank setting, where a lexical ranker initially retrieves a subset of candidate passages, which are then reordered more accurately by a dense ranker.

In this thesis, we focus on the task of passage re-ranking.
We employ a dual-encoder architecture as re-ranker that employs a two independent query and document encoders, allowing document embeddings to be pre-computed. Dense query-passage similarity is computed as a dot product between their representations.
We then combine scores from both stages using score interpolation.

We identify query encoding latency as a bottleneck and propose an Average Embedding (AvgEmb) estimator. This novel model can efficiently predict an accurate query representation, without requiring any attention-based encoding.
It operates solely on looking up embeddings and computing their weighted average representation.
Our model is distilled from a TCT-ColBERT and achieves 98.6% of its teacher's accuracy while being 13.4X more efficient in query latency and 1.6X better in the full interpolated passage re-ranking pipeline on CPU.

Our code is publicly available on https://github.com/BovdBerg/fast-forward-indexes. ... Read more

Accurate Wind Resource Assessment (WRA) requires the correction of systematic errors in modeled wind fields using sparse and temporally limited on-site measurements. With the emergence of microscale Large-Eddy Simulation (LES) as a high-resolution alternative to traditional mesoscale models, current correction practices, such as Measure-Correlate-Predict (MCP) for temporal extrapolation and Inverse Distance Weighting (IDW) for spatial extrapolation, need to be re-evaluated and adapted to the microscale context. This thesis investigates how data-driven methods can improve LES wind-speed error correction in both time and space, addressing three research gaps: the lack of validation of MCP on microscale data, the absence of temporal and spatial context in standard MCP-style formulations, and the limited use of flow information in spatial extrapolation.

The temporal component evaluates a set of linear and nonlinear regression models as flexible MCP-style baselines and extends them with additional contextual information. Temporal structure is incorporated through multi-step time windows, and local spatial structure is added by including neighboring LES grid cells. To represent full-field atmospheric patterns, the models are further enriched with latent encodings of the LES wind-speed field obtained through a Convolutional Autoencoder. The results show that MCP generalizes well to the microscale and that adding temporal and spatial context improves time-series accuracy across all observation locations, with combined strategies outperforming industry-standard methods.

The spatial component introduces Wind Speed-enhanced IDW (WS-IDW), which augments traditional IDW by weighting observation locations not only by geographic distance but also by similarity in LES wind speed. WS-IDW produces consistent improvements over the baseline, particularly when more observation locations are available. Analysis of the correction maps reveals that WS-IDW partially smooths misplaced fine-scale streaks in the LES wind field, supporting the hypothesis that LES is prone to slight spatial misalignment of coherent structures. The proposed method generalizes reasonably across sites and across different numbers of source masts.

Together, the temporal and spatial results demonstrate that incorporating LES-derived spatial and temporal information yields systematic improvements in microscale correction performance. The thesis provides a refined understanding of how LES behavior interacts with data-driven correction methods and offers a foundation for developing more robust microscale WRA correction frameworks in future work. ... Read more

Mirages are a visual phenomenon consisting of the appearance of a mirrored image of an object, without the presence of an actual mirror surface, due to light rays that are curved because of continuous refraction in the air, which relates to differences between the surface and ambient temperature.
As temperature is defined in space, while a typical standard rasterization pipeline only processes surfaces, these phenomena are difficult to reproduce.
Approximating the nonlinear ray path with ray marching becomes taxing due to the long light paths.
Current approaches use acceleration structures and have not been implemented in a rasterizer.

We present two methods that make dynamic real-time rendering of mirages possible, which fit well in the rasterization pipeline.
The first solution uses a second camera to capture surface temperature and normal information below the view ray, and approximates the nonlinear path of the ray in as few steps as possible.
The second method obtains the surface information in screen space instead, making it faster, but potentially less accurate in heterogeneous scenes.

Results show that both methods are capable of rendering mirages dynamically and in real-time when the surface is relatively flat.
Therefore, both methods, especially the second, faster method, could be used for the rendering of mirages on relatively flat faces, enabling real-time dynamic rendering of mirages in video games on those types of surfaces. ... Read more

Music recommender systems are increasingly present in our lives, and it is important to keep trying to improve recommendations in order to make them match the users preferences as well as possible. To achieve this, a vast amount of song and user data has to be analysed and taken into account. One of the approaches to do this, includes analyzing different audio features in order to find other songs with similar traits. The majority of the research and data in this sector is focused around adults, with little research surrounding children, which can result in worse recommendations for this demographic. In this paper, the focus is shifted towards children with the purpose of filling that gap. This is achieved by examining the prominence of specific song features among children of different age groups, expanding the knowledge on listening habits of a major demographic. More specifically, the research presented in this paper explores the prominence of various song features, aiming to find a connections between these features and the listening habits of children of specific age ranges from 8-18. This paper’s conclusions will offer potential enhancements, which can improve existing recommender systems by considering findings for their design. These findings will therefore allow for a more tailored experience for children of different age ranges, increasing overall user experience. ... Read more

Recommender systems play a considerable role in the consumption of music, also for children. Children are easily influenced, inappropriate song lyrics can negatively impact children's behaviour and personality, by teaching inappropriate language or harmful biases. We argue for a need for recommenders to protect children from inappropriate content. Recommender systems should consider what content is and is not well-suited for children. To guide future design and development of recommender systems for children, this research probes the music lyrics of Pop, Rock, Country, Rap, and R&B for inappropriate content. We achieve this by an empirical analysis using four existing algorithms to scan different facets of inappropriate content in the lyrics in a dataset of 37,993 songs from the Genius database. The outcome of this empirical exploration reveals that Rap and R&B are the most inappropriate music genres and advises to be cautious when recommending these genres to children. Results also show that inappropriate content is highly prevalent in music. We address the need for a filter for recommender systems, to filter inappropriate songs for children. ... Read more

Recommender systems play a large role on contemporary music platforms, but they tend to work less well for non-mainstream listeners such as children. Additionally, there is no one strategy to perfectly capture a listener's music preference. As children develop understanding of music in different stages, using features they respond to might make recommendations more accurate. Therefore, this study seeks to investigate the effectiveness of a recommender which utilises matrix factorisation augmented with the musical features of tempo, mode, dynamics and time signature in recommending songs a child user would like. We evaluate the quality of this recommender based on the Factorisation Machine algorithm by comparing it to a non-augmented variant of the same algorithm and similar ones using fewer of the same features. Results show that while adding features improves the quality of recommendations, adding too many or the wrong features diminishes said improvement, although more research is needed in this direction. ... Read more