This publication presents an established engineering-driven framework for pipelining large language model (LLM) agents for automatic code translation. The framework is applied in a long-term ongoing industrial project with Danfoss Power Solutions aimed at translating five million lines of Delphi to C#. To manage this scale, the source codebase is divided into translatable chunks, which are processed independently through an agentic LLM pipeline, with subsequently the reassembly of the translated code chunks in the target language. Due to the inherent stochasticity of LLMs, the framework incorporates processing and validation functions to ensure consistency and correctness. As an extended version of our short paper (Bruneel et al., 2025), this paper provides a significantly more comprehensive architectural blueprint. Specifically, we present an indepth background & related work section, formalise the chunking and reassembly strategies, detail the internal mechanics of agent stages, and introduce sequential agent orchestration alongside the dynamic LLM pipeline architecture. Furthermore, we report early empirical observations regarding the pipeline's computational overhead, validation dynamics, and translation quality, noting a practical impact of an initial 10x acceleration over estimated manual translation efforts. Ultimately, we share these expanded insights to inform future development and application in similar large-scale translation tasks.

Many real-world modeling problems are hindered by limited data availability. In such cases, transfer learning leverages related source domains to improve predictions in a target domain of interest. We extend the classical gradient tree boosting paradigm to a regression transfer algorithm by modeling the weak learner as a sum of two regression trees. The trees are fitted on source data and target data, respectively, and jointly optimized for the target data. We derive optimal coefficients for the model update under the least-squares, the least-absolute-deviation, and the Huber loss functions. We benchmark our approach against boosting-based regression transfer methods in twelve transfer scenarios. The results indicate that our approach constitutes a competitive alternative within the realm of boosting-based regression transfer. Moreover, we provide a theoretical justification as well as empirical validation that our approach is robust under larger domain shifts.

A software system's source code is organized into files, but the way those files group into architectural modules is often undocumented or has drifted from the original design. Software architecture recovery (SAR) reconstructs this module view to help developers understand and maintain large systems.

We study SAR from file-level static dependency graphs by learning file embeddings in a self-supervised manner and clustering them into modules. Using node2vec as a baseline, we train graph autoencoders via link reconstruction and compare dependency-based representations (a graph attention encoder, with and without personalized PageRank diffusion) against heterogeneous ones that also incorporate the project's folder structure. We evaluate eleven open-source Java, C, and C++ systems against ground-truth architectures using complementary alignment and modularity metrics. Recovery from a dependency structure alone is useful but limited, bounded more by how much structural context is available to the model than by the choice of encoder. Folder organization helps where dependency cohesion is weak, but heterogeneous methods often gain alignment by drifting from pure dependency modularity, a trade-off between structural cohesion and the developer-intended decomposition. No single method wins on every system and metric — node2vec and the diffusion model are comparable on dependencies alone, while folder-aware methods are worth their added complexity mainly when cohesion is weak. All representations scale well and are practical to run and tune, so representation learning is not the bottleneck for adoption.

Remote sensing techniques are widely used for mapping and monitoring forest attributes, providing valuable information on forest cover, biomass, and overall forest health. In recent years, national airborne laser scanning (ALS) campaigns have been conducted in several countries to map forest resources. When combining ALS with field inventory data, these datasets enable the development of nationwide models for prediction of forest attributes. In this study, we explore the potential of machine learning (ML) to enhance existing modeling approaches for nationwide forest attribute mapping in Sweden. We achieve this by relating ALS data from the most recent ALS campaign of Sweden with field data from the Swedish National Forest Inventory (NFI). By aggregating laser metrics from surveyed areas (NFI plots), as well as over surrounding areas to the plots, we investigate (1) if ML approaches can outperform existing linear regression baseline models and (2) if further enhancements of the predictive capacity can be achieved by including surrounding, spatially correlated ALS data. To this end, we used extreme gradient boosting (XGBoost), as well as a convolutional neural network (CNN), specialized to handle tabular data and spatially correlated data, respectively. The models were evaluated on five forest variables: basal-area weighted mean tree height, basal-area weighted mean stem diameter, basal area, stem volume, and above-ground biomass. All models were evaluated on several nested datasets to assess the robustness, showcasing consistent results across datasets. We achieved significant improvements in prediction accuracy across all investigated forest variables. Furthermore, incorporating surrounding information to the modeling rendered further improvements for diameter, basal area, and biomass predictions. The approaches tested and developed here thus form a promising basis for flexible modeling approaches that can be transferred globally for large-scale forest monitoring and management.

The detection and localization of quality-related problems in industrially mass-produced products has historically relied on manual inspection, which is costly and error-prone. Machine learning has the potential to replace manual handling. As such, the desire is to facilitate an unsupervised (or self-supervised) approach, as it is often impossible to specify all conceivable defects ahead of time. A plethora of prior works have demonstrated the aptitude of common reconstruction-, embedding-, and synthesis-based methods in laboratory settings. However, in practice, we observe that most methods do not handle low data quality well or exude low robustness in unfavorable, but typical real-world settings. For practitioners it may be very difficult to identify the actual underlying problem when such methods underperform. Worse, often-reported metrics (e.g., AUROC) are rarely suitable in practice and may give misleading results. In our setting, we attempt to identify subtle anomalies on the surface of blasted forged metal parts, using rather low-quality RGB imagery only, which is a common industrial setting. We specifically evaluate two types of state-of-the-art models that allow us to identify and improve quality issues in production data, without having to obtain new data. Our contribution is to provide guardrails for practitioners that allow them to identify problems related to, e.g., (lack of) robustness or invariance, in either the chosen model or the data reliably in similar scenarios. Furthermore, we exemplify common pitfalls in and shortcomings of likelihood-based approaches and outline a framework for proper empirical risk estimation that is more suitable for real-world scenarios.

Mapping source code entities manually to architectural modules is labor-intensive and time-consuming. Automating this process can help adopt static architecture compliance checking techniques like reflection modeling. We propose a graph-based semi-supervised learning approach to automate this mapping. Our method leverages syntactic and semantic dependencies extracted from compiled software artifacts and formulates the problem as a node classification task on a graph. We investigated Graph Convolutional Network (GCN) variants: (1) a homogeneous graph approach that treats all dependencies uniformly and (2) a heterogeneous graph approach that maintains distinct dependency types, utilizing a Relational Graph Convolutional Network (RGCN). The proposed method was evaluated on eight open-source Java systems and benchmarked against the state-of-the-art Naive Bayes classifier. Results demonstrate that the RGCNN variant outperforms the baseline, offering improved accuracy for mapping source code entities to architectural modules.

This paper presents two experiments testing the capabilities of the models GPT-4-vision-preview and GPT-4o, to transcribe images of randomly generated number-strings. We noticed a tendency for longer number strings to be read inaccurately while transcribing invoices using the models, so we created an experiment with images with minimal noise to see what amount of digits these models are able to transcribe with full accuracy. We further tested weather the mistakes that occur in transcribing longer strings are reoccurring when the images are re-tested. We analyzed the character of the mistakes made, and weather any digits are over represented among the ones involved in mistakes. We found, among more results described in the paper, that the models are 100% accurate up to 75 digits per image, that the same mistakes are reoccurring when the same image is rerun, and that hallucinations are only 23% of all mistakes made by the models.

Pixel-wise time-series analysis of satellite-derived vegetation indices is often complex and computationally demanding, as the pixel-wise application of the analysis scales with the number of pixels to be processed. In this study, we use Gaussian Process Regression (GPR) pixel-wise to fill temporal gaps in Normalized Difference Vegetation Index (NDVI) data. Rather than applying GPR across all pixels and trying to optimise the performance of GPR directly, we introduce an algorithm to select key pixels for GPR applications. By selectively applying GPR to a subset of pixels and image reconstruction of the remaining unselected pixels, we achieve substantial speedups (2.5-5.5×) with minimal loss in NDVI estimation accuracy (0.5-2.5%). This trade-off with pixel selection enables more scalable and efficient pixel-wise processing of large-scale imagery, with strong applicability in the field of remote sensing.

This paper provides details and analysis of findings from three experiments on the capabilities of the OpenAI chatbot ChatGPT-4 with vision capabilities as a dual-purpose tool for Optical Character Recognition (OCR) and key-value data extraction tasks. Our results could be relevant to any task where one is interested in extracting key information from images with approximately equal complexity, and are scalable to be used for large datasets. We did the main experiments in the OpenAI user interface for the model, and a smaller experiment using the API. The experiment used a dataset comprising 1000 digital invoices alongside 1000 photographic images of real receipts, collected from a broad spectrum of market sectors within Sweden. The main experiment gave us a significant accuracy rate, achieving $\mathbf{9 9. 8 \%}$ in extracting critical financial information from the digital invoices. Similarly, when applied to the photographic images of receipts, it maintained a high accuracy level of $\mathbf{9 9. 5 \%}$. The smaller experiment gave us an accuracy of $\mathbf{9 4. 4 \%}$ These findings are particularly noteworthy not only because of the high accuracy levels but also due to the model's effectiveness in performing both OCR and key-value extraction tasks as a one-step process. This dual functionality underscores the model's potential as a highly efficient and reliable solution in automating financial data extraction and processing tasks.

The appearance of the log cross-section provides important information when assessing the quality of the log, where properties to consider include pith location and density of annual rings. This makes tasks like estimation of pith location and annual ring detection of great interest. However, creating labeled training data for these tasks can be time-consuming and subject to misjudgments. For this reason, we aim to create generated training data with controlled properties of pith location and amount of annual rings. We propose a two-step generator based on generative adversarial networks in which we can completely avoid manual labeling, not only when generating training data but also during training of the generator itself. This opens up the possibility to train the generator on other types of log end data without the need to manually label new training data. The same method is used to create two generated training datasets; one of entire log ends and one of patches of log ends. To evaluate how the generated data compares to real data, we train two deep learning models to perform estimation of pith location and ring counting, respectively. The models are trained separately on real and generated data and evaluated on real data only. The results show that the performance of both estimation of pith location and ring counting can be improved by replacing real training data with larger sets of generated training data.