This paper presents a novel approach to geolocalization, a task that aims to predict the latitude and longitude of an image based on its visual content. Traditional methods in this domain often rely on databases, complex pipelines or large-scale image classification networks. In contrast, we propose a direct regression approach that simplifies the process by predicting the geographic coordinates directly from the image features. We leverage a pre-trained Vision Transformer(ViT) model, specifically the ViT of a pre-trained CLIP model, for feature extraction and introduce a regression head for coordinate prediction. Various configurations, including pre-training and task-specific adaptations, are tested and evaluated resulting in our model called ReGeo. Experimental results show that ReGeo offers competitive performance compared to the existing approaches, despite being simpler and with minimal supporting code. ReGeo consistently achieves the best performance at the continent level across the GWS15K, IM2GPS3K, and IM2GPS datasets, and even outperforms GeoCLIP at the country level on GWS15K. Notably, GWS15K exhibits a significant distribution shift compared to ReGeo's training data, highlighting the model's strong generalization capabilities. On the holdout dataset, ReGeo localizes 32 % of 5,000 images at the city level, and achieves increasingly higher accuracy at broader radii—88.08% at the regional level, 96.86 % at the country level, and 99.34 % at the continent level-demonstrating its effectiveness for mid-range to coarse-grained geo-localization tasks.
Despite the proliferation of Responsible Artificial Intelligence (RAI) principles, organizations struggle to translate them into practical implementation. This study investigates the challenges Swiss organizations face in implementing RAI through qualitative interviews with industry professionals and academic experts, complemented by a multi-stakeholder workshop. We first identify five critical pain points hampering RAI implementation: economic constraints, structural and procedural barriers, conceptual and technical challenges, cultural and behavioral resistance, and regulatory uncertainty. Then we propose the Control-Tangibility Framework, a novel framework that maps pain points along two fundamental dimensions: organizational control and challenge tangibility. Our framework provides organizations with a structured methodology to prioritize RAI efforts by considering both their ability to influence change and their capacity to observe aspects of the challenges. Furthermore, we provide practical insights for developing targeted implementation strategies that bridge the gap between ethical principles and operational practices. Our findings suggest that successful RAI implementation requires moving beyond compliance-focused approaches toward a comprehensive organizational transformation, supported by systematic assessment and prioritization of implementation challenges.
Large Language Models (LLMs) have demonstrated capabilities beyond basic text generation, like question answering, translation, and even stylistic text generation. Since these models are available for public use, enormous effort has been put into safety engineering to ensure that undesirable and harmful text is not generated and the generations are polite and empathetic. In this work, we examine the inherent empathy capabilities of five open-source LLMs and evaluate them from multiple angles using automated metrics to understand their capabilities and limitations. In the context of this work, “inherent” refers to the LLM's ability to generate empathetic text without having to explicitly prompt for it. We examine if empathy is treated as a style change or is the model demonstrating some understanding of the specific user's context. We find that LLMs use more emotion words than humans in their generations. They can also infer the user's emotional state, a crucial characteristic of empathy. Due to the probabilistic nature of obtaining generations, there is a tendency for the responses to drift away from the user's actual intent. In such cases, specific prompting allows the model to respond appropriately. We summarize the differences observed between human and LLM generations and conclude with a potential research direction for empathetic dialog generation that leverages the capabilities of LLMs.
Automated Essay Scoring (AES) is a challenging task requiring models to evaluate writing quality with human-like consistency while being computationally efficient and scalable. Current approaches often rely on computationally intensive prompt engineering or resource-heavy transformer models, limiting their practical application. In this study, we present a novel prompt-free approach using SetFit for AES that achieves competitive accuracy while significantly reducing computational overhead. Unlike traditional transformer- based models such as DeBERTa, SetFit enables sentence transformer fine-tuning with contrastive learning, making it suitable for essay scoring even in low-data regimes. To handle full-length essays, we extended Longformer into a custom sentence transformer with a 4096-token context window, overcoming the 512-token limitation of standard transformer architectures. Our fine-tuned model, published on Hugging Face, has gained over 6,000 downloads, indicating strong community interest in efficient, prompt-free approaches. Our results demonstrate that SetFit with an extended Longformer sentence transformer achieves competitive performance, making it a viable alternative to computationally expensive models. Beyond essay scoring, our approach shows promise for other long-form text analysis tasks such as legal document review, research paper assessment, and content quality evaluation in educational contexts. This work contributes to the growing exploration of efficient NLP methods for educational assessment, offering a practical alternative to resource-intensive prompt-based solutions.
In the era of large language models (LLMs), the detection of software vulnerabilities in code snippets remains a critical challenge, particularly in resource-constrained environments. This paper explores the efficacy of fine-tuned and zero-shot LLMs, and traditional machine learning (ML) models for classifying code snippets as "vulnerable" or "not vulnerable" based on the Software Common Weakness Enumeration (CWE) framework. We perform an evaluation of three fine-tuned LLMs: Flan-T5-small, Llama-3.2-1B-Instruct and Deepseek-llm7B-Base-alongside three traditional ML models: Logistic Regression, Multinomial Naive Bayes, and Linear Support Vector Classifier. We then compare these approaches with zero-shot models - Deepseek-R1 and ChatGPT-o1. Our experiments show that LLMs achieve competitive validation and test accuracies, with Llama-3.2-1B-Instruct and Deepseek-llm-7B-Base both achieving 84% accuracy on the test set. However, traditional ML models, particularly Logistic Regression and Linear SVC, demonstrate strong performance with 78% accuracy, highlighting their viability in scenarios where computational resources are limited. Furthermore, we address the challenges of dataset imbalance, resource-intensive fine-tuning, and the practical applicability of these models in real-world vulnerability detection pipelines. Our study emphasizes the importance of considering traditional ML models as viable alternatives to LLMs, especially in resource-constrained environments while contributing to the ongoing research on the practical applicability of AI, and data science in software security.
As Large Language Models continue to grow in size and sophistication, more benchmarks are needed to assess their capabilities. Benchmarks are crucial for measuring progress and pushing the scientific community forward. However, they are often expensive and exploitative to build. In this paper, we demonstrate how to translate the rich and complex existing resources for SQL and relational databases that have been used by the database community, for SPARQL and Knowledge Graph Question Answering (KGQA). We describe how we bootstrapped the largest and most complex KGQA benchmark to date, explore the complexities of converting a relational database into a knowledge graph and describe the methods used to learn rich schema information in the KG from its relational counterpart using a query workload-based analysis. In addition, we show how this method can be used to bootstrap any existing SQL benchmark for KGQA by translating the newly created ScienceBenchmark dataset into a KGQA dataset.
AI algorithms for the automatic detection of unusual or abnormal patterns in image data have become increasingly important in industrial quality inspection, improving product quality and operational efficiency. Most state-of-the-art Image Anomaly Detection (IAD) methods are based on unsupervised approaches, learning normal patterns from anomaly-free training data. However, in real-world applications the assumption of anomaly-free training data is often unrealistic, as labeling anomalies in the historical data can be expensive, error-prone, or even impossible. Anomalies contaminating the training data typically lead to a degraded anomaly detection (AD) performance at deployment, yet this issue remains largely overlooked in research. Some studies have attempted to mitigate this challenge through data refinement methods. However, these approaches often require prior knowledge of the anomaly ratio (AR) in the training data, which is rarely available in practice. In this paper, we introduce Overlapping Subsets Data Refinement (OSDR), a simple, fully unsupervised, and model-agnostic refinement framework designed to address image anomaly detection (IAD) under data contamination with no prior assumptions about the AR. OSDR assigns a refinement score to each training sample using an ensemble of models trained on partially overlapping data subsets, followed by robust anomaly removal through an adaptive thresholding technique. Evaluations on two widely used industrial image datasets demonstrate that OSDR effectively restores performance losses caused by contamination and outperforms existing refinement frameworks. Our approach provides a flexible, practical, and easy-to-deploy solution for IAD in real-world settings, where data contamination or mislabeling is often inevitable and the anomaly ratio is unknown.
In Switzerland, over 100 metric tons of plastic and other foreign materials are introduced into agricultural fields through the application of contaminated compost fertilizers every year. Annually, 154 kg of biogenic waste is generated in Switzerland per capita, and it often contains contaminants such as food packaging, coffee capsules, and plastic bags. Although a significant portion of these contaminants is removed during processing, small fragments (,20 mm) persist in the final digestate or compost fertilizers. Currently, quality control relies on visual inspection of dried and sieved 2 kg samples, two to twelve times a year, and per site. This represents only a negligible fraction of the 1.26 million metric tons of biogenic waste produced annually. Testing a larger amount would increase the significance of the quality control. But the current manual approach is costly. An automated approach to the quality control of compost fertilizer according to legal requirements is needed to increase the sample size. In this study, we demonstrate the feasibility of detecting the most common plastic contaminants in compost fertilizers, down to a few millimeters in size, using advanced imaging technologies in the visible and near-infrared domain, combined with state-of-the-art machine learning algorithms. We collected and curated our own datasets, consisting of images of compost fertilizer or digestate, contaminated with pieces of plastic foil or shards of hard plastic, mostly polypropylene (PP) and polyethylene (PE). These datasets were used to train multiple segmentation models. Resulting in a retrieval rate of over 75% of all objects, ranging from multiple centimeters to just a few millimeters. The error of the total predicted area lies at just 2% compared to the ground truth, showing that our method can be applied to predict the contamination level. The results of our RGB-camera-based model show the potential to automate and significantly accelerate the quality check procedure with cost-effective hardware, allowing testing of a larger amount of material. Our solution could be applied for continued quality control during production, by adding our system to the end of compost fertilizer production lines on already existing conveyor belts.
Automated tissue segmentation in medical imaging plays a critical role in clinical AI-assisted decision-making, and particularly in the assessment of body composition from CT scans. However, acquiring data and annotations of sufficient quality to train deep-learning models is expensive and timeconsuming. In this work, we propose a novel approach to improve data efficiency and model accuracy by leveraging domain knowledge about biologically relevant tissue-specific Hounsfield unit (HU) ranges as an inductive bias for learning. Specifically, we extend the input representation of deep learning-based segmentation models with binary masks indicating potential tissue types, where each binary mask is created from thresholds derived from medical literature. Our method not only enhances segmentation performance by up to 5% for intramuscular adipose tissue but surpasses the performance of the baseline model with 50% of the training data. Our easy-to-apply method thus improves data efficiency and facilitates the development and use of segmentation models in resource-constrained clinical settings.
We developed a data analytics tool in which snow and weather data as well as historical avalanche accidents are analyzed to compute avalanche risks. The focus of the tool is the monitoring of the current situation in the Swiss Alps to provide an overview and to warn people before starting their walking, hiking, or skiing activities in the mountains. To reach this goal we applied dimensionality reduction and clustering techniques to project the data to lower dimensions and to structure it. The avalanche accident data is additionally included to allow the judgment of the current and past snow weather situations in terms of how risky it is for people when doing their daily mountain activities. We illustrate the usefulness of the data analytics tool by applying it to data measured at several hundred measurement stations at different locations and elevations in the Swiss Alps. Finally, we discuss scalability issues and limitations before we conclude with future work.
Online platforms are popular among legal pro fessionals for daily searches of information needed for drafting opinions, preparing court cases etc. These platforms must handle large volumes of documents ensuring timely availability of information. A key challenge is identifying and linking citations in the documents to their sources within the limited time available for nightly re-indexing. Traditional methods, such as regular expressions, are fast but lack flexibility and do not support context-dependent citations. The latest state-of-the-art generative language models lack efficiency for the task, processing only 36 documents per hour. We propose transformer-based encoder-only models to recognize citations. We demonstrate that the discriminative BERT-based models can process 3,000 documents in about 32 minutes, exceeding the performance requirements and ensuring scalability and efficiency of large-scale legal platforms.
Coding of veterinary electronic health records (vEHRs) using standardized terminologies, such as Systemized Nomenclature of Medicine - Clinical Terms (SNOMED-CT), promises to be a powerful addition to modern clinical research allowing for interoperability and efficient information retrieval and analysis. The process of manually coding vEHRs is timeconsuming and expensive, creating a need for automated approaches. This study investigates targeted pre-training and fine-tuning of significantly more compact model architectures to develop a competitive model at a fraction of the computational cost required for state-of-the-art vEHR coding tools. Three sizes of transformerbased LLMs and two sizes of Mamba LLMs were pre-trained on a corpus of 636,754 veterinary visit summaries and fine-tuned on 199,642 labeled visit summaries. Competitive accuracies were achieved with RamGPTM and RamGPTL models at a fraction of the computational cost while fitting on commodity laptops. Selected Mamba models did not perform as strongly as selected transformer-based models and did not exhibit appreciably faster pre-training and inference times.
Dental X-rays play a crucial role in assessing oral health and supporting the diagnosis of conditions affecting teeth and bone structures. We investigated the automated identification of areas of interest for potential caries lesion candidates in Xrays to support the human caries detection process. To achieve this, we investigated whether computationally efficient semantic segmentation models can rival the performance of instance segmentation models. We challenged a highly accurate two-stage instance segmentation approach by comparing it with semantic segmentation enhanced with advanced training strategies that use unlabeled data and tooth detection information. Compared to most previous work, we trained our models on a larger dataset (4,000 bitewings). Our top semantic model, the multitask Dynamic U-Net (F1 0.69, recall 0.66, precision 0.73), scored comparably to our Mask R-CNN instance model (F1 0.65, recall 0.80, precision 0.54). Both models achieved scores comparable to or higher than those of previous studies that used object detection metrics and larger datasets. Qualitative expert evaluation confirmed that semantic segmentation models perform similar to instance segmentation. Mask R-CNN training took two to three times longer than for Dynamic U-Net and DeepLabV3+, respectively. These findings suggest that semantic segmentation models provide a computationally efficient automation approach to visualize areas of interest for caries lesion candidates.
Continual learning (CL) mandates the capacity to learn novel tasks sequentially while preserving knowledge from previously encountered ones. However, deep networks trained strictly incrementally often suffer from catastrophic forgetting, where updates for new tasks overwrite representations learned from earlier tasks. In this work, we introduce Meta-Replay with Adaptive Feature Fusion (MRAFF), a unified framework that addresses catastrophic forgetting while maintaining scalable and memory-efficient updates. First, MRAFF employs dynamic network expansion by adding minimal task-specific blocks or heads, ensuring model capacity grows only when essential for new tasks without disrupting prior knowledge. Second, it introduces a feature-fusion autoencoder for latent replay, storing compact latent representations and labels rather than raw data, thereby reducing storage overhead. These stored embeddings are selectively replayed based on a meta-replay strategy that estimates the utility of reintroducing prior tasks. Crucially, this meta-replay mechanism controls when and how many past representations to reconstruct, striking a balance between maintaining older skills and adapting to new ones. We validate MRAFF on Split MNIST, showing that it retains substantial accuracy on earlier tasks with negligible additional memory cost, outperforming standard replay-based baselines. By compressing data into a latent space and coupling replay with meta-learned sampling, MRAFF emphasizes a practical path toward real-world CL scenarios requiring adaptability and efficient resource usage.
The last decades have seen alternating focus on two approaches to natural language processing: (a) mapping the sequence of past words or tokens to the forecast of the next word, represented by the Attention Mechanism (AM), and (b) recurrent processing of the sequence while keeping contextual information in the memory. The quadratic complexity of AM sets limits to the sequence lengths. This has led to the revival of interest in recurrent models, now in the form of State Space Models (SSMs). SSMs are a traditional approach from systems and control theory. Expanding SSMs into the convolutional form shows, after some transformations, essential structural similarity to attention processing. Existing proofs of formal similarity do not point out some aspects of functional difference. This work applies the system-theoretical view to assign both approaches to fundamentally different model categories. It is argued that AM corresponds to the parameter-extensive concept of Finite Impulse Response models, while SSMs are related to the parametereconomical Infinite Impulse Response models. The second, more compact, category has a more general modeling potential. This finding motivates the search after resolving the inability of SSMs to explicitly capture the relevance of the context independently of the position. Based on this, an extension of SSMs is proposed, which introduces the concept of similarity, which is analogous to that of the attention mechanism. Implementing and testing this extension is nontrivial and will be the task for the future.
Despite the growing challenge of product returns in e-commerce, limited research explores real-time prediction of return intent. Existing studies typically rely on customer relationship management (CRM) data and require historical context, such as a user's return behavior the last n orders. However, many online shoppers lack this context, because they either browse as guests and can thus not be identified in realtime or are first-time visitors. This reduces the accuracy and coverage of existing return intent assessment methods, which would otherwise lay the groundwork for timely interventions. This paper explores the potential for return prediction based only on browsing behavior using Long Short-Term Memory (LSTM) models. We simulate real-world settings by training and testing on a rolling-window approach on logged browsing data. We find for our model to attain F1 = 0.68 and ROC AUC = 0.70 on average. Furthermore, our model demonstrates superior accuracy compared to methods relying on CRM data without historical user context. A subsequent analysis of the important features, based on the trained LSTM model, draws insights into which browsing behaviors may be early warning signals indicating return. For instance, having a high number of items in the shopping cart from a previous browsing session is strongly associated with high-value returns. Meanwhile, for desktop users, the use of filters appears to be a driving indicator for returns. In a hypothetical calculation, we illustrate the positive economic impact our approach enables to e-commerce companies. Our findings highlight the potential of utilizing browsing behavior for prediction of return intent, providing a foundation for developing real-time return intent assessment strategies.
This paper presents an explainable decision support system (DSS) for stress classification in virtual reality (VR)-based tele-rehabilitation, addressing the critical need for transparency in machine learning (ML) applications within healthcare. It leverages SHapley Additive explanations (SHAP) to estimate feature relevance, enabling therapists without technical expertise to validate stress detection outcomes. By providing contextualized feature importance, the system bridges the gap between complex ML models and domain-specific knowledge, enhancing usability. This user-centered approach promotes informed decision-making and advances the integration of explainable AI in clinical environments.
The integration of multimodal data, combining both textual and visual information in a single context, has become increasingly important in many applications involving transformers. However, most existing models primarily focus on multimodal input, generating only textual outputs. This paper presents the Multimodal Autoregressive U-Net Transformer (MUT), a novel architecture designed to enable true multimodality by processing and generating both textual and visual tokens within the same context. By leveraging autoregressive large language models (LLMs) and incorporating a U-Net-inspired encoder-decoder for multiscale image processing, MUT seamlessly integrates and generates multimodal outputs. Experiments on the COCO2017 dataset demonstrate the model's ability to segment and reconstruct images based on textual prompts without requiring postprocessing steps. Despite some limitations in handling out-ofdomain images, the MUT architecture shows significant potential for tasks that require close integration of visual and textual modalities.
As energy systems evolve toward greater integration of renewables and distributed generation, understanding and modeling consumption patterns becomes increasingly important. Traditional clustering techniques often produce average profiles and can fail to capture the complex variability within private customer data. Generative diffusion models address this limitation by producing realistic, diverse synthetic consumption profiles while preserving confidentiality. In this paper, we present our initial work on diffusion-based profile generation to anonymize real customers' smart meter data. We demonstrate that diffusion models generate more representative synthetic profiles than other existing approaches. Finally, we discuss how conditioning on external variables-for example, photovoltaic (PV) installation or electric vehicles-can add new capabilities for tailored profile generation. We outline these conditional approaches as next steps, emphasizing their potential to further enhance privacy and realism in future energy consumption scenarios.
