DeLTA 2026 Abstracts

Area 1 - Computer Vision Applications

Full Papers

Paper Nr:	26
Title:	Task-Aware Detection of Elongated Objects in Wide-Field Industrial Imagery
Authors:	Matous Cejnek and Martin Vitoušek
Abstract:	Detecting elongated objects in wide-field industrial imagery presents challenges that are not well addressed by general-purpose object detection frameworks. In production environments such as steel processing lines, objects often appear as long, thin structures spanning a large portion of the image width, while visual conditions may include reflections, scale, and partial occlusions. Accurate localization along the transport axis is therefore essential for reliable downstream automation. We present a lightweight dense detection framework designed for extreme aspect-ratio imagery. The proposed model follows an anchor-free formulation and combines a compact depthwise-separable backbone with a feature pyramid network. To better capture long-range dependencies along the dominant object orientation, we introduce a directional context module implemented using large horizontal depthwise convolutions. Object instances are recovered from dense predictions through peak detection followed by lightweight non-maximum suppression. The method is evaluated on an industrial dataset containing elongated objects observed in wide-field production scenes. Across repeated train–validation splits, the proposed model achieves detection performance comparable to modern YOLO-based detectors while requiring substantially fewer parameters and lower computational cost. These results suggest that incorporating simple geometric priors into the architecture can improve efficiency for structured industrial detection tasks.

Paper Nr:	30
Title:	DT-KAD: Dual-Teacher Knowledge and Attention Distillation for Efficient Breast Cancer Diagnosis using Ultrasound Imaging Data
Authors:	Pratiksha Gawas and Sowmya Kamath S
Abstract:	Knowledge distillation (KD) has emerged as a powerful method for training compact models with limited data and computational resources for medical image analysis. The recent works in KD goes beyond the logit-based supervision to incorporate multi-view and multi-teacher learning. However, many existing KD-based approaches remain limited in how they distill spatial information and use complementary knowledge from heterogeneous teachers, particularly in limited-data settings. In this work, we propose the dual-teacher knowledge and attention distillation (DT-KAD) framework that integrates logit-based supervision with deep, spatially aligned feature map (attention) distillation from two complementary teacher networks. The distilled student model is deliberately designed to be lightweight, with a small size, making it suitable for resource-constrained clinical settings. It has very few parameters (203 283) and has tiny size of 794.07 KB with training GFlops of 0.493. We trained and tested our DT-KAD approach on the small BUSI dataset for breast cancer classification. DT-KAD with the proposed lightweight student model performed decently well on limited data achieving an weighted Accuracy, Precision, Recall, F1-score, CKS and MCC of 0.856, 0.856, 0.856, 0.853, 0.746, 0.7493 respectively, is 99.08% lighter, 11.56 times faster, and has 99.11% fewer parameters when compared to the best teacher model. The proposed approach promotes spatially informed representations, mitigates representation collapse during distillation, and improves classification performance and interpretability. Experimental validation underscored the superiority of the proposed approach in limited data and resource settings.

Paper Nr:	51
Title:	Camera-to-Subject Distance: An Overlooked Quality Component in Mobile Face Biometrics
Authors:	Arto Apila, Dorina Rajanen and Jukka Komulainen
Abstract:	The accuracy of facial verification systems depends on image quality which in turn is influenced by capture conditions among other factors. While ISO/IEC 29794-5:2025 defines standardized face image quality components, it does not explicitly account for magnification distortion caused by short camera-to-subject distances (CSD), which are common in mobile and self-enrollment scenarios. This limitation is particularly relevant for border control applications such as the Entry/Exit System (EES), where facial biometrics are increasingly captured under unconstrained and mobile-assisted conditions. In this work, we present a systematic empirical analysis of the relationship between CSD, magnification distortion, standardized face image quality scores, and biometric verification performance. The study uses multiple datasets captured at varying distances, including a controlled experimental dataset, a public multi-subject dataset, and a distortion-corrected dataset. Facial image quality is assessed using the OFIQ toolbox, and verification accuracy is evaluated with multiple face recognition models. The results show that increasing magnification distortion consistently degrades face verification performance across multiple recognition models and datasets, despite facial images achieving high standardized quality scores. This indicates that current face image quality metrics do not sufficiently account for capture-related geometric distortions induced by camera-to-subject distance present during acquisition. Based on these findings, we propose a CSD-aware enrollment and verification workflow that integrates distortion awareness into existing biometric pipelines to improve robustness in mobile and border control deployments.

Short Papers

Paper Nr:	17
Title:	PointConViT: An Adaptable Convolution-Transformer Model for High-rate 3D Detection in Autonomous Robot
Authors:	Ammar Naich, Nikesh Bajaj, Sofia Zahri and Jesús Requena Carrión
Abstract:	Fast and reliable 3D object detection is essential for autonomous vehicle (AV) navigation and real-time robotic perception. Although vision transformers have shown promise for 3D point cloud detection, their use remains challenging due to the sparsity, irregularity, and large scale of point cloud data. Hybrid CNN-Transformer models partially address these issues, but often encounter low detection rates and difficult parameter tuning. To overcome these limitations, we proposes PointConViT, an end-to-end transformer model designed for highrate 3D object detection. Unlike hybrid architectures, PointConViT uses a pure transformer backbone with a Gated Positional Self-Attention (GPSA) mechanism, enabling both local feature aggregation and global contextual reasoning. By allowing early transformer layers to behave similarly to convolutional layers, PointConViT learns local and global 3D features within a unified architecture. The model is evaluated on the KITTI 3D Object Detection Benchmark against state-of-the-art methods. For the Car class, PointConViT achieves a 3D AP of 84.98% on the Easy setting and a BEV AP of 81.55% on the Hard setting, while reaching the highest detection rate of 15.6 FPS. These results show that PointConViT provides a strong balance between detection accuracy and computational efficiency, making it suitable for real-time 3D object detection on point clouds in AV. Our code is publicly available at https://github.com/ammaryasirnaich/pointconvit.

Paper Nr:	29
Title:	Evaluating SAM3 and MedGemma for Brain Tumor MRI: Zero-Shot Segmentation, Linear Probing, and Multimodal Diagnosis
Authors:	Tamara Kostova, Ivan Kitanovski, Monika Simjanoska Misheva and Kostadin Mishev
Abstract:	This paper presents a series of experiments applying two recent foundation models—SAM3 [30], a text-prompted image/video segmentation model, and MedGemma 1.5 4B [31], a multimodal medical vision–language model—to brain tumor analysis tasks. We first evaluate SAM3 in a zero-shot setting on all 125 patients from the BraTS 2021 dataset [8], observing limited performance (Dice = 0.189, IoU = 0.124, Sensitivity = 0.397). To address this, we apply linear probing: a single linear layer is trained on frozen SAM3 image-encoder features, substantially improving segmentation quality to Dice = 0.836 (pixel-level aggregate) and 0.801 (per-case mean, 95% CI [0.770, 0.832]), a gain of +0.647 over the zero-shot baseline without modifying the backbone. Independently, we evaluate MedGemma 1.5 4B on a multi-dataset brain MRI benchmark [10, 15, 12, 11] covering five classification tasks: imaging modality, MRI sequence, image plane, binary diagnosis, and tumor subtype. Finally, we introduce a combined pipeline in which SAM3 provides spatial attention by overlaying a detected bounding box onto the full image before passing it to MedGemma. The pipeline improves tumor detection accuracy (85.1%→96.3%) and image-plane prediction (28.2%→51.6%), but introduces a critical normal-scan specificity collapse caused by SAM3 false-positive detections (67.1%→41.3%).

Paper Nr:	48
Title:	A Priori Capacity Estimation for Encoder-Decoder Architectures in Unpaired CAD-to-SEM Image Translation
Authors:	Marcelo Bacher
Abstract:	Selecting optimal model capacity for image-to-image translation networks typically requires extensive hyperparameter search. We present a capacity estimation framework for encoder-decoder architectures in unpaired CAD-to-SEM translation for semiconductor manufacturing. The framework combines five information-theoretic factors—image size, domain gap, frequency mismatch, texture complexity, and rate-distortion ratio—to bound the bottleneck dimensionality before training; aggregation weights are empirically calibrated, making the estimator a theory-guided structured heuristic rather than a purely analytical predictor. A novel application of Radial Power Spectral Density (R-PSD) analysis quantifies the frequency-domain gap between multichannel CAD and SEM images, yielding stable R-PSD estimates across one synthetic and three real industrial datasets. On synthetic data, the predicted optimal configuration achieves an SSIM of 0.625 and an LPIPS of 0.632, with under- and overfitting configurations showing 13–17% SSIM degradation and 1–5% LPIPS degradation. Compared with stateof-the-art methods, the proposed approach achieves the best LPIPS (0.632) with competitive SSIM (0.625), while requiring 30% fewer inference-time parameters than U-GAT-IT and MUNIT, suggesting that principled capacity estimation may offer a practical path towards more resource-efficient deployment in semiconductor inspection applications.

Paper Nr:	50
Title:	Spatiotemporal Anomaly Detection in Surveillance Videos Using CNN–TCN Fusion with Multiple Instance Learning
Authors:	Aditya Singh Yadav and Abubakr Siddig
Abstract:	The extensive deployment of surveillance cameras has increased the demand for automated systems capable of detecting anomalous human behaviour in video streams, as manual monitoring is inefficient and prone to error. This study addresses anomaly detection in untrimmed surveillance videos, with emphasis on maintaining accuracy, computational efficiency, and robustness under real-world constraints. The research examines how spatio-temporal information can be effectively integrated using weak supervision. Existing approaches often rely solely on spatial or temporal features and frequently employ computationally intensive architectures unsuitable for deployment. To overcome these limitations, the paper proposes a lightweight fusion architecture combining a Convolutional Neural Network (CNN) and a Temporal Convolutional Network (TCN) within a Multiple Instance Learning (MIL) framework. Spatial features are extracted using a ResNet-50 backbone, while temporal dependencies are modelled through dilated convolutions in the TCN. Keyframe selection is applied to reduce redundancy while preserving semantic information. Experiments conducted on the UCF-Crime dataset demonstrate that CNN-only models favour precision, whereas TCN-only models better capture temporal context but suffer from higher false positive rates. The fused model balances these strengths, achieving a more balanced trade-off between precision and recall, despite a slightly lower AUC of 88.2% than the CNN baseline. The findings highlight the effectiveness of lightweight spatio-temporal fusion for reliable and deployable video anomaly detection systems.

Paper Nr:	52
Title:	Infrared as a Fallback: A Multimodal Defense Approach against Adversarial Attacks on RGB Object Detection
Authors:	Yeshwanth Kumar Adimoolam and Christos Kyrkou
Abstract:	Single-modality object detectors deployed in safety-critical autonomous systems are vulnerable to adversarial perturbations in the visible spectrum. Even modest pixel-level noise or carefully optimized patches can collapse the precision and recall of state-of-the-art detectors, causing significant issues for downstream planners and operators. We present a multimodal detection pipeline that exploits the complementary failure modes of RGB and infrared (IR) sensors to provide a graceful, low-cost fallback whenever the visible modality is compromised. The pipeline runs two independently trained YOLO detectors, one per modality, and fuses their outputs through an asymmetric IoU-based merging rule that allows the IR stream to override or supplement RGB detections under high spatial overlap. We instantiate the pipeline using YOLOv3 on the LLVIP low-light pedestrian dataset and YOLOv10 on the DroneVehicle aerial dataset, and test it with two adversarial attacks: additive Gaussian noise as a sensor-degradation surrogate, and the Cross-Modal Patch Attack (CMPA) as an adaptive cross-modality threat. Across both datasets, the fused pipeline preserves a substantial fraction of clean detection performance under Gaussian noise and cross modal patch attacks where the visible stream undergoes significant degradation of its baseline mAP, while the infrared is able to retain a higher percentage of the clean detection performance. Our results position infrared imaging as a practical and sensor-level defense layer against adversarial attacks that are commonly targeted towards RGB sensors.

Paper Nr:	27
Title:	Image-Based Prediction of Apple Harvest Timing
Authors:	Martin Vitoušek, Matouš Cejnek, Jan Vrba, Jakub Jura and Lucie Plecitá
Abstract:	Accurate estimation of fruit maturity is essential for planning harvest operations in commercial orchards. While computer vision methods have been widely applied to fruit detection and yield estimation, predicting harvest timing directly from images remains a challenging task due to biological variability and seasonal effects. In this work, we investigate the problem of estimating the number of days relative to the optimal harvest date from single images of apple tree canopies. We evaluate several modern convolutional neural network backbones and analyze their performance under two evaluation protocols: stratified cross-validation and temporal cross-validation, which separates training and testing by collection day. Experiments are conducted on a dataset spanning multiple days before harvest for a single apple cultivar. Across 31 folds, the evaluated models achieve test mean absolute errors between 7.8 and 9.0 days. While validation performance varies across architectures, the final test errors remain relatively similar, suggesting that model capacity has a limited impact on predictive accuracy in this task. The results further highlight substantial variability across folds, indicating that environmental conditions and temporal differences in data collection play a significant role in prediction difficulty. These findings suggest that future improvements may require incorporating additional temporal or environmental information beyond single-image inputs.

Paper Nr:	43
Title:	Assessing Cross-Task Consistency for Adversarial Robustness in Visual Perception
Authors:	Rafaella Elia, Christos Kyrkou and Theo Theocharides
Abstract:	Multi-task consistency provides a useful signal for detecting adversarial manipulation in visual perception pipelines that jointly perform object detection and instance segmentation. Under normal conditions, predictions from these tasks are naturally aligned, for example, detected object regions tend to overlap with semantically compatible segmentation areas. Adversarial perturbations, even when designed to degrade a single task, often disturb this alignment because both heads depend on shared evidence and frequently share feature representations. Building on this observation, this work replicates and extends a lightweight, training-free detection approach that quantifies agreement between detection and segmentation outputs via a task-consistency score and flags inputs that deviate from the clean consistency distribution. The method is evaluated under gradient-inspired adversarial perturbations analogous to FGSM and PGD, across multiple perturbation magnitudes, reporting both consistency trends and detection performance using threshold-based metrics and ROC--AUC analysis. Experiments on two benchmarks with distinct scene characteristics show that adversarial examples induce measurable consistency shifts in cross-task agreement, enabling reliable attack detection without retraining or modifying the underlying models. These findings highlight multi-task consistency as a practical and interpretable mechanism for robustness monitoring in modern multi-head perception systems.

Paper Nr:	57
Title:	Structure-Aware Line Order Estimation for Classical Japanese Documents
Authors:	Tetsuya Suzuki and Asahi Sakurai
Abstract:	Estimating the reading order of text lines is a fundamental problem in document image analysis, particularly for classical Japanese documents containing complex and irregular layouts. Conventional rulebased approaches based on simple geometric heuristics often struggle to handle annotations, wrapped lines, and non-uniform line arrangements. In this paper, we propose a structure-aware and layout-adaptive framework for reading order estimation in classical Japanese document images. The proposed method models pairwise relationships between text lines using text-line centerline images while preserving their structural shape information. A deep neural network predicts the relative reading order between pairs of text lines, and the final reading sequence is estimated by aggregating pairwise predictions. The proposed model is trained using synthetically generated document images designed to simulate the layout characteristics of the target document. Experiments using real document images derived from a facsimile of The Tale of Ise demonstrate that the proposed method can successfully estimate reading order even for complex layouts where simple rule-based methods fail. The best model achieved a Kendall’s τ of 0.9904. These results suggest that document specific learning using synthesized layout data is a promising strategy for reading order estimation in classical Japanese documents.

Area 2 - Models and Algorithms

Full Papers

Paper Nr:	32
Title:	DeLIVeR: Decomposed Learning for Information-Grounded Veracity Recognition via Reinforced Knowledge Graph Exploration
Authors:	Cong Hoan Nguyen, Thomas Hoang, Hieu Minh Duong and Long Nguyen
Abstract:	Automated fact-checking remains a challenge for Large Language Models (LLMs) due to "query brittleness" in traditional retrieval systems. We propose DeLIVeR (Decomposed Learning for Informationgrounded Veacity Recognition), a framework that treats evidence retrieval as a reinforced strategic exploration task. DeLIVeR utilizes a Planner LLM to decompose complex claims into targeted question sets, which are used to traverse structured Knowledge Graphs (KGs) for highprecision evidence. We optimize the Planner’s policy using Group Relative Policy Optimization (GRPO) with a reward system prioritizing structural diversity and verdict accuracy. Our evaluation on LIAR, FEVER, and PolitiFact shows that DeLIVeR significantly outperforms state-of-the-art baselines. Using Qwen2.5-7B, our framework achieved peak F1-scores of 83.73, 84.57, and 79.70 respectively, representing a 10–15% improvement over HippoRAG2. By shifting to a reinforced question-planning strategy, DeLIVeR effectively bridges multi-hop reasoning gaps and provides an auditable, transparent path for verifiable misinformation detection.

Paper Nr:	34
Title:	Long-Document Prediction under Label Scarcity: Comparing Transformer Fine-Tuning, Parameter-Efficient Tuning, and Classical Baselines
Authors:	Pinar Ersoy and Mustafa Erşahin
Abstract:	Long-document prediction under label scarcity is a difficult problem at the intersection of document intelligence, low-resource learning, and efficient adaptation. In legal, scientific, financial, and regulatory settings, predictive evidence is often distributed across long textual spans, while annotation remains costly and domain-expert dependent. This study presents a comparative evaluation of three methodological families for low-label long-document prediction: classical sparse-feature baselines, full transformer fine-tuning, and parameter-efficient tuning. A unified protocol is defined over three long-context corpora with class-balanced label budgets of 8, 16, 32, and 64 examples per class. The compared methods include TF-IDF with linear classifiers, frozen embedding baselines, full fine-tuning of Longformer and BigBird, and parameter-efficient variants based on adapters and LoRA. The results show that full fine-tuning is not uniformly optimal under severe supervision scarcity. Parameter-efficient tuning achieves the best balance between macro-F1, calibration, and adaptation cost at the lowest budgets, while classical baselines remain competitive when labels correlate strongly with lexical regularities. The findings indicate that adaptation granularity is a primary modeling variable in low-label long-document learning rather than a secondary engineering choice.

Paper Nr:	39
Title:	When Alignment Feels Artificial: Evaluating Preference-Based Fine-Tuning for Perceived Humanness
Authors:	Rebecca Ansell and Lisa Singh
Abstract:	We investigate whether fine-tuning on human preference data can enhance perceived humanness in AI-generated political social media text. Using pairwise judgments comparing human- and AI-authored posts about the 2024 U.S. presidential debate, we fine-tune models through Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO). While both methods successfully modulate stylistic properties, increasing emotional expressiveness and lexical diversity, these changes do not translate to improved perceived humanness. All model variants were rated as nearly indistinguishable, with preference-optimized models showing no advantage over baselines. Inter-annotator agreement was near chance, revealing high variance in how evaluators perceive authorship. We suggest that standard fine-tuning methods may optimize toward a central tendency, potentially reducing the natural stylistic diversity that characterizes human expression—particularly in short-form social media text, where stylistic variation is high. Our findings suggest that enhancing perceived humanness may require methods that preserve or expand output variance rather than converging on an averaged style.

Short Papers

Paper Nr:	25
Title:	Fairness-Aware Generative Adversarial Imputation for Longitudinal Disease Subtype Discovery in Parkinson’s Disease
Authors:	Moad Hani, Fatima Zahra Ouardirhi, Mohamed Amine Hamzaoui, Zainab Ouardirhi, Saïd Mahmoudi and Mohammed Benjelloun
Abstract:	This work investigates whether fairness-aware imputation can act as an enabling condition for reliable, demographically equitable longitudinal disease-subtype discovery in Parkinson’s disease (PD). Using the Parkinson’s Progression Markers Initiative (PPMI) cohort, we explicitly target age- and gender-related disparities by stratifying the population into three age groups (younger than 60, 60-70, older than 70) and by gender, and quantify imputation bias along these axes prior to any subtype inference. Six imputation methods (LMM, MICE, HyperImpute, GAIN, MissForest, mean-imputation baseline) are jointly evaluated with classical accuracy metrics (MAE, RMSE, R2) and three fairness metrics adopted from the algorithmic-fairness literature (SPD, DI, CDD). Generative Adversarial Imputation Networks (GAIN) achieve the best joint accuracy-fairness trade-off (MAE=0.30 ± 0.09, 88% bias reduction over the baseline). Fairness-validated, GAIN-imputed trajectories are then passed to the Subtype and Stage Inference (SuStaIn) algorithm, which discovers six clinically distinct subtypes with progression rates ranging from 1.2 to 2.9 points/year. External validation on 387 AMP-PD participants shows excellent agreement (ICC=0.85); subtype-specific prognostic models reach AUROC 0.73-0.81 versus 0.68 for a non-stratified baseline. The complete pipeline-data preprocessing, GAIN training, fair- ness evaluation, SuStaIn inference and figure generation-is available from the corresponding author upon reasonable request, together with the dependency specifications required for end-to-end reproduction; requests are answered within 15 business days.

Paper Nr:	15
Title:	A Decision Pipeline for RAG Architecture Selection in Medical Education: Empirical Analysis and Cost-Performance Guidelines
Authors:	Jovana Dobreva, Tadej Horvat, Matjaz Gams, Igor Mishkovski, Kostadin Mishev and Monika Simjanoska Misheva
Abstract:	We propose a structured decision pipeline to support practitioners in selecting Retrieval-Augmented Generation (RAG) architectures for medical education tasks, and we report a preliminary empirical comparison that informs the pipeline’s recommendations. Four RAG variants, Vanilla RAG, Vector RAG, Graph RAG, and Hybrid RAG, were evaluated on the 420 multiple-choice questions of the RAGCareQA dataset, and we report accuracy with 95% Wilson confidence intervals, end-to-end inference time, and estimated per-query cost derived from published token pricing. Overall accuracies cluster tightly between 96.25% and 96.50% with overlapping confidence intervals, and pairwise Fisher’s exact tests do not reveal statistically significant differences between architectures on the full set or on any question-type subset. Despite the absence of significant overall differences, descriptive results suggest plausible architecture-task affinities: Graph RAG was correct on all 23 graph-enhanced questions (95% Wilson CI [85.7%, 100%]) at roughly 3× the estimated cost and 4-8 s latency, while Vanilla RAG matched the field on simpler questions at baseline cost and 1-2 s latency. We use these descriptive patterns, rather than statistical separation, to motivate a three-phase decision pipeline (Requirements Analysis, Constraint Evaluation, Architecture Assessment) that helps practitioners navigate costperformance trade-offs. We position the pipeline as a proposed framework that requires further validation on larger and more diverse datasets, and we identify the absence of a no-RAG baseline, single-judge automated scoring, and the small size of the graph-enhanced subset as the principal limitations of this preliminary study.

Paper Nr:	53
Title:	An Enhanced Machine Learning Approach for Cryptocurrency Price Prediction Using Hybrid Metaheuristic Optimizers
Authors:	Aykhan Bandaliyev, Turkana Gahramanli and Ahmed M. Khedr
Abstract:	The volatile nature of cryptocurrencies requires meticulous analysis of their indicators before investing or trading with them. This study aims to develop a framework that predicts devaluation in cryptocurrency values using machine learning and feature selection based on a hybrid PSO (Particle Swarm Optimization)-GA (Genetic Algorithm) algorithm. Four supervised machine learning models-SVM (Support Vector Machine), Logistic Regression, Decision Trees, and Random Forest are applied on a time-series data with a total of 10 features. After receiving the initial results from execution of these four algorithms, the PSO metaheuristic algorithm is implemented for feature selection. First, eight features are selected, then, the search space is expanded to select between six and nine features. After the PSO implementation, a hybrid PSO-GA (Genetic algorithm) is applied in feature selection instead of simple PSO to find a search space with more feasible candidates to refine results. As a result, the SVM model based on the hybrid PSO-GA algorithm feature selection outperformed others.

Area 3 - Natural Language Understanding

Full Papers

Paper Nr:	22
Title:	Selective Invariance and Uncertainty under Evidence-Presentation Shift: Toward Robust Multi-Hop Reasoning for Verification and Adversarial Evidence Settings
Authors:	Ningning Liu, Vimala Balakrishnan and Zhuangzhuang Pan
Abstract:	Multi-hop question answering (QA) systems are typically trained and validated under one evidence presentation but deployed under another. A common mismatch arises when the same retrieved evidence is later shown as a long, noisy concatenation of documents, or as a compressed context constrained by latency and memory. This paper studies evidence-presentation shift in a controlled setting with paired inputs derived from the same retrieval result. Two motivations of the present study are: (i) when complete evidence is present but formatted differently, answer beliefs should remain stable; (ii) when a required reasoning hop is missing or corrupted, models should reduce confidence rather than remain over-certain. We construct complementary evidence views around an anchor presentation, including an answer-preserving sufficient view and two answer-breaking views that drop or swap a supporting hop. Experiments on HotpotQA and 2WikiMultiHopQA show strong directional asymmetry in baseline transfer. Targeted view construction yields large gains in full-to-packed transfer, where compact evidence is easier to consume, while producing more modest and sometimes mixed effects in the harder packed-to-full direction. Our approach improves several cross-regime robustness metrics and explicitly trains models to reduce confidence under corrupted evidence, highlighting the importance of calibrated uncertainty in settings such as misinformation, rumor propagation, fact checking, and online verification.

Paper Nr:	23
Title:	Simple Token-Efficient Vision-Language Model for Case-Level Pathology Synoptic Report Generation
Authors:	Zhiyuan Yang, Jiahao Cheng, Vincent Quoc-Huy Trinh and Mahdi S. Hosseini
Abstract:	Generating clinically useful pathology reports for pathology cases from whole-slide images (WSIs) is challenging due to gigapixel resolution, long visual-token sequences, and the complexity of case-level reasoning, where a single case may contain multiple WSIs with heterogeneous tissues and ambiguous findings. We present a simple token-efficient vision–language model for case-level synoptic report generation that remains practical under constrained GPU memory. Our architecture follows a minimal three-component design: a frozen pathology patch encoder, a lightweight two-layer MLP vision-language aligner, and a large language model decoder, with an explicit WSI marker token to separate slides within a case. Training proceeds in two supervised stages: (1) aligner-only WSI captioning using heterogeneous WSI-text pairs, and (2) case-level supervised fine-tuning on case-report pairs for structured report generation. To reduce sequence length, we represent each slide using 512 × 512 patches at 5× magnification, which reduces the average sequence length by up to 64× times compared to the commonly used 20× patches. Combined with efficient training techniques, we enable practical training with only half a NVIDIA H100 GPU. Across both training stages, our approach achieves high ROUGE-L/METEOR/BLEU-4 scores while being substantially more efficient in memory and runtime. In AI-based evaluations, our model is consistently preferred over strong baselines. Extensive ablations characterize performance-efficiency tradeoffs and identify simple choices that improve robustness in multi-WSI settings. Overall, this work provides a strong, reproducible baseline for efficient pathology report generation, lowering the barrier to multi-WSI VLM research under limited compute.

Paper Nr:	56
Title:	Do Deployment Constraints Make LLMs Hallucinate Citations? An Empirical Study across Four Models and Five Prompting Regimes
Authors:	Chen Zhao, Yuan Tang and Yitian Qian
Abstract:	Large language models (LLMs) are increasingly used to draft scholarly text and to support evidence synthesis, but they often hallucinate bibliographic references that look legitimate. We contribute a structured evaluation framework for citation reliability under deployment constraints: a three-way label scheme (Existing/Unresolved/Fabricated), a deterministic scoring pipeline against Crossref and Semantic Scholar, human-audit calibration that characterizes the framework’s biases, and a released defensive tool (citecheck). Applying the framework to 144 claims across six domains, we evaluate two proprietary models (Claude Sonnet, GPT-4o) and two open-weight models (LLaMA 3.1–8B, Qwen 2.5–14B) across five regimes: Baseline, Temporal (publication-year window), Survey-style breadth, Non-Disclosure policy, and their combination. In the closed-book five-regime study, across 17,443 generated citations, no model exceeds a citation-level existence rate of 0.475; Temporal produces the steepest single-condition decline, and the stacked Combo condition yields the lowest overall verifiability, while outputs remain formatcompliant (well-formed bibliographic fields). Unresolved outcomes constitute a large share of outputs (36–61%); in a 200-citation stratified human audit, 44 of 87 pipeline-Unresolved citations (50.6%) were judged fabricated. To probe whether explicit grounding improves closed-book citation verification, we add a minimal retrieval-augmented baseline for Claude Sonnet that injects top-5 candidate records retrieved from Crossref (the DOI registration agency) into the same prompts; this raises Baseline existence from 0.381 to 0.699 (∆ = +0.318) and Temporal existence from 0.119 to 0.818 (∆ = +0.699). For these two cells, grounding substantially improves verifiability, though this minimal diagnostic baseline is not a full four-model RAG comparison. Because the framework has known label uncertainty, our conclusions rest on the relative pattern of effects across conditions and model families rather than on absolute fabrication rates. The pipeline is released alongside the paper to support post-hoc citation verification before LLM outputs enter scholarly literature, journalism, or policy briefings.

Short Papers

Paper Nr:	33
Title:	Understanding the Voice of the Customer in the Airline Industry: A Comparative Study of Traditional ML, Fine-Tuned BERT, and Fine-Tuned GPT-4o-Mini for Supervised Topic Classification
Authors:	Aylanur Cüce
Abstract:	This paper presents a fully supervised comparative study of airline customer feedback topic classification across three modelling paradigms: a traditional TF-IDF + Logistic Regression baseline, a fine-tuned BERT encoder, and a fine-tuned GPT-4o-mini generative classifier. The task is formulated as a nine-class operational classification problem using approximately 600,000 English-language customer feedback records collected from digital channels. Unlike weakly supervised, unsupervised, or zero-shot approaches, the study uses an existing operational subject field as the supervisory label space and evaluates all models under the same taxonomy and the same train, validation, and test split. The evaluation combines aggregate metrics, class-wise F1-scores, confidence intervals, pairwise statistical testing, error-overlap analysis, and operational benchmarking. The results show that the fine-tuned BERT model achieves the best overall performance, while the traditional baseline remains competitive and operationally efficient. The fine-tuned generative model improves substantially after output normalization, indicating that generative classification can be viable in closed-label enterprise settings when supported by output governance. Overall, the findings suggest that contextual encoder models provide the strongest performance for business-defined airline feedback routing, while traditional machine learning remains attractive when efficiency and deployment simplicity are critical.

Area 4 - Machine Learning

Full Papers

Paper Nr:	13
Title:	Parameter-Efficient Adaptation of Multilingual ASR Models to Swiss German Dialects
Authors:	Marija Stojcheva, Goran Petkovski and Igor Mishkovski
Abstract:	Automatic Speech Recognition (ASR) for low-resource and dialect-rich languages remains challenging due to phonological and lexical variation, the absence of standardized orthography, and the limited availability of consistently annotated speech data. Although recent multilingual and foundation ASR models demonstrate strong zero-shot performance across many languages, their accuracy often degrades when applied to dialect-rich and low-resource speech, particularly under pronounced dialectal variation and domain mismatch. Swiss German exemplifies these challenges, as its dialectal diversity and diglossic relationship with Standard German limit reliable ASR performance. This paper presents a systematic evaluation of state-of-the-art ASR systems, including multiple Whisper checkpoints, a decoder-only Voxtral model, and GPT-4o-based transcription, across four Swiss German datasets (ASGDTS, SPC, SwissDial, and STT4SG) covering diverse dialect regions and recording conditions. Beyond zero-shot analysis, parameter-efficient adaptation using Low-Rank Adaptation (LoRA) is investigated to fine-tune a Whisper Large-v3 Turbo model, and a hybrid architecture is introduced that integrates the adapted Whisper encoder into a Voxtral-style decoder-only ASR framework. Experimental results show that parameter-efficient fine-tuning consistently improves performance across all datasets, reducing word error rate by up to 50\% relative to strong zero-shot baselines while updating fewer than 1\% of model parameters. The LoRA-adapted Whisper model achieves the best overall performance and establishes state-of-the-art results on SwissDial and STT4SG, while the hybrid Whisper–Voxtral system remains competitive. These findings demonstrate that lightweight adaptation provides an effective and computationally efficient solution for robust ASR in low-resource and dialectally diverse settings.

Short Papers

Paper Nr:	18
Title:	Rethinking Model Evaluation for Urban PM2.5 Classification: Imbalance, Temporal Validation and Computational Cost
Authors:	Ziwei Shu, Jesús Cáceres-Tello, Orlando Lima Rua and Ramón Alberto Carrasco
Abstract:	Model performance in urban air quality prediction is commonly summarized through aggregate metrics that reward average behaviour. However, the most consequential events in practice are rare high-pollution episodes. This raises a critical issue: can evaluation choices alter which model appears optimal? This work analyzes daily PM2.5 classification in Lisbon using official open data from the European Environment Agency covering the period from 2021 to 2023. A strictly prospective temporal validation scheme is adopted, with models trained on 2021–2022 observations and evaluated on an unseen 2023 test set to replicate op-erational conditions. Four algorithmic families are compared: Logistic Regres-sion, Random Forest, XGBoost, and a Multi-Layer Perceptron (MLP). Evalua-tion prioritizes imbalance-aware metrics, notably Macro F1 and Balanced Accu-racy, and incorporates computational cost, defined as the combined training and inference time, as an explicit comparison dimension. When temporal integrity and class imbalance are properly accounted for, model rankings shift in meaningful ways. The MLP achieves the most stable cross-class performance and substan-tially higher recall for high-pollution episodes, while simpler models remain competitive under strict computational constraints. The findings show that model superiority in urban environmental systems is contingent on evaluation design. Evaluation design therefore plays a decisive role in determining model suitability for real-world deployment, with implications for how environmental Artificial In-teligence systems are benchmarked and implemented.

Paper Nr:	20
Title:	Strategic Absenteeism Profiling: Integrating Fuzzy Linguistic RFM and Explainable AI in People Analytics
Authors:	Ziwei Shu, Álvaro Carrasco-Aguilar, Mario Arias Oliva and Orlando Lima Rua
Abstract:	Workforce absenteeism poses a significant challenge for organizations, affecting both productivity and financial performance. Traditional Human Resources (HR) approaches often rely on uniform corrective measures that overlook the diversity of absence behaviors. To address this gap, this paper aims to develop and validate a novel People Analytics framework for the strategic segmentation of employee absenteeism. The framework is based on an adaptation of the marketing-derived Recency, Frequency, and Monetary (RFM) model. Following the CRISP-DM methodology, the proposed approach combines K-means clustering with the 2-tuple linguistic model, integrating absence patterns (Recency and Frequency) with employee salary data (Monetary) to construct a multidimensional risk profile. The analysis identifies four distinct workforce segments: Low-Cost Chronic Absen-teeism, Stable High-Value Talent, Emerging Escalating Risk, and Sudden High-Impact Alert. Crucially, to ensure the robustness of these unsupervised profiles, the framework incorporates an Explainable Artificial Intelligence (XAI) valida-tion phase. A surrogate modeling strategy using Random Forest and SHapley Additive exPlanations (SHAP) is employed to mathematically validate and inter-pret the feature contributions driving each cluster. Generative AI is also leveraged to translate these complex segmentations into actionable executive insights. By merging advanced soft computing techniques with HR strategy, the proposed framework enables precise risk profiling, supports personalized talent manage-ment strategies, and informs proactive organizational interventions.

Paper Nr:	35
Title:	Trustworthy Multimodal XAI in Customs: A Segmentation-Based Study
Authors:	Otmane Amel, Xavier Siebert and Sidi Ahmed Mahmoudi
Abstract:	Multimodal models that combine different data types, such as images and text, have shown strong results in many classification tasks. In customs, where declarations can include product images along with textual descriptions, such models can help to improve goods classification and risk assessment. When these models are used in production, it becomes important to explain their predictions, especially for the visual part, where explanations are harder to interpret. Visual explanations often depend on how an image is split into regions. If the regions do not match real objects (or meaningful parts), the explanation can be noisy and hard to trust. In this work, we study how the choice of segmentation affects post-hoc explanations for multimodal Harmonized System (HS) code classification in the context of a customs fraud detection solution. Using a fixed multimodal classifier, we generate explanations using attribution-based post-hoc methods, and for the visual modality, we compare a traditional superpixel segmentation with a foundation-model-based object-aware segmentation to define the interpretable regions. We evaluate its faithfulness using the deletion area under the curve (DAUC) metric and also analyze the visual coherence of the produced heatmaps. Experiments on real customs declarations show that a good segmentation choice strongly improves visual explanations. Foundation-model-based regions achieve a mean image DAUC of 0.031, a significant improvement over 0.044 with traditional superpixels, and also produce more coherent heatmaps, while text attributions remain stable. These results suggest that context-aware segmentation is a key component when visual explanations are used to support customs inspection decisions.

Paper Nr:	45
Title:	AI Framework for Role-Aware Predictive Decision Support in Artillery Operations Grounded in Multimodal Biomonitoring
Authors:	Michal Sustr, Tomáš Havlík, Daniel Korec, Martin Vlkovský, Jan Drozd, Luděk Rak, Jiří Néma, Vilém Hanuš and Milan Brabec
Abstract:	Modern artillery operations expose personnel to tightly coupled cognitive, physical, and environmental stressors that degrade performance before observable failure. Existing wearable sensing and machine learning approaches remain limited by treating operators as homogeneous and by lacking actionable command-level outputs. This paper addresses this gap by proposing a role-conditioned predictive framework linking multimodal biomonitoring with decision support in artillery operations. The framework estimates short-term performance degradation risk across roles with fundamentally different workload profiles, such as fire direction center personnel and weapon system crews, and transforms individual predictions into team-level readiness and mission-level capability indicators. It is designed for deployment in denied, degraded, and disrupted operational environments (D3OE), emphasizing local inference, robustness to incomplete data, and controlled operation under uncertainty. Rather than presenting a fully validated system, the paper defines a structured foundation for integrating human-state prediction into command decision processes, shifting biomonitoring from descriptive assessment toward predictive and actionable decision support.

Paper Nr:	47
Title:	Multi-Lens Triple Extraction with Composed Loss
Authors:	Adrian-Zoltan Herczeg and Mihaela Dinșoreanu
Abstract:	With the current evolution of the natural language processing domain and the state-of-the-art models found at the pinnacles of research, knowledge representation under graph forms has solidified itself as one of the strongest techniques. At the core of this method, we find the triple extraction process. In the current landscape, the task is performed alongside others by very robust models, which, although achieving very impressive performances, are expensive to train in both time and computational effort. We explore the possibility of a lightweight solution for triple extraction that tries to strike a balance between performance and computational power, while keeping the underlying structure simple. Specifically, we introduce a strategy we call multi-lens triple extraction, which explores the possibility of training a lightweight model using multiple focal losses calibrated specifically for each component of a triple to maximize the performance of identifying each component. The results show a constant increase in performance between 0.02 and 0.03 points in F1 score compared to using a single loss function.

Paper Nr:	49
Title:	Horizon-Adaptive Differenced Switching
Authors:	Michel Spils and Sven Tomforde
Abstract:	Recent time series forecasting models often achieve strong average accuracy, but their performance can vary substantially across the forecast horizon. In particular, some architectures produce stable early predictions while deteriorating more strongly at longer horizons. This paper proposes Horizon-Adaptive Differenced Switching (HADiS), a simple post-processing strategy that combines standard forecasts with differenced forecasts depending on their horizon-specific validation performance. The method identifies local minima in the direct forecast error and switches to differenced predictions when they promise improved accuracy. We evaluate HADiS on several standard multivariate forecasting benchmarks and with different model architectures, including Transformer- and MLP-based approaches. Our results show that HADiS can improve forecast accuracy in a model- and dataset-dependent manner, especially when direct forecasts exhibit pronounced horizon-wise instability. The proposed method is lightweight, model-agnostic, and easy to integrate into existing forecasting pipelines.

Paper Nr:	54
Title:	Enhancing GAN-Based Video Summarization via Diffusion Score Refinement
Authors:	Berkay Sancı and Yusuf Yaslan
Abstract:	Video summarization aims to produce a concise visual representation of a full-length video, ensuring the flow of the content and the most important segments are preserved. Most of the unsupervised approaches utilizing Generative Adversarial Networks (GANs) have shown success by formulating the summarization task as a video reconstruction problem. Recent methods have introduced Self-Attention mechanisms for frame selection, replacing LSTMs, improving the capture of long-range temporal dependencies between video frames. However, the importance scores generated by self-attention may exhibit high variance and temporal inconsistency, which degrades the quality of frame selection. In this work, we propose a novel framework that integrates a Diffusion-based Score Refinement mechanism into the adversarial training loop. Specifically, we employ a Denoising Diffusion Probabilistic Model (DDPM) to act as a regularizer for the output of the self-attention module. During training, the diffusion process regulates the noisy attention scores, enforcing importance scores that are representative. This refined scores guide the Variational Autoencoder (VAE) and GAN network to reconstruct the video using only the most important segments. This diffusion-driven regularization allows our model to learn robust scoring policies without requiring computationally expensive iterative sampling during inference. Experimental results show that the proposed method outperforms traditional GAN-based methods in capturing key segments and maintaining the integrity of the summary.

Paper Nr:	24
Title:	Sequential Recommendation with Multi-View Diffusion
Authors:	Shin Ando
Abstract:	Leveraging multiple behavior types to model users' interactions and preferences is a critical challenge for recommendation systems. We propose Multiview Diffusion (MvDiff), which captures the patterns of interactions from multiple views, or subsequences, of a user history. It consists of a series of blocks that extracts different sets of behaviors from the views forwarded from the previous blocks to generate views with progressively increasing ratios of noisy behaviors (pageview). Denoising for each view is conditioned on a reconstructed sequence from another view, so as to capture the relations between types of behaviors, e.g., target (purchase), interest-related (add-to-cart, add-to-favorite), by the denoising and conditioning networks and reconstruct user behaviors in small, incremental steps. MvDiff benefits from performing diffusion in a continuous latent space and leveraging multi-view conditioning. The proposed model addresses denoising as a series of incremental, smaller generative subtasks, avoids embedding for item-behavior combinations, and reduces sensitivity to noisy behaviors by performing denoising diffusion conditioned across multiple views. Our empirical study shows that MvDiff achieves competitive performance and improved robustness with noisy interactions against existing models.

Paper Nr:	36
Title:	Benchmarking AI Models for Early Detection of Elevated Load States from Multimodal Wearable Data toward Artillery Readiness Assessment
Authors:	Martin Blaha, Natálie Gottvaldová, František Krompolc, Alena Oulehlova, Ivana Nekvapilová, Jan Drozd, Luděk Rak and Milan Brabec
Abstract:	Wearable sensing is increasingly used for stress and health monitoring, yet many approaches remain difficult to translate into operational decision support because they focus on isolated physiological states rather than deployment-relevant load as-sessment. In readiness-oriented military applications, an important early methodolog-ical question is which AI model families, sensing configurations, and feature repre-sentations are most suitable for robust detection of elevated load states from multi-modal wearable signals. This paper addresses that question through a benchmark-oriented design using public multimodal wearable datasets as structured surrogate environments. Rather than claiming direct prediction of artillery combat readiness, the study formulates subject-independent elevated-load detection as a proxy task rel-evant to later readiness-oriented analytics. The reported experiments focus on an ini-tial WESAD-derived binary benchmark, comparing logistic regression, linear SVM, and random forest models together with full, compact HRV-oriented, and frequency-only feature representations. Initial results on a WESAD-derived benchmark indicate that interpretable linear models, especially logistic regression, provide the strongest balance between detection performance, robustness, and practical deployability. In addition, a compact cardiovascular/autonomic feature set centered on heart-rate and HRV-related descriptors retains substantial predictive value. The findings should therefore be interpreted as evidence for the feasibility of a first binary warning layer, not as validation of a complete readiness-assessment system. These findings support a conservative development path in which readiness-oriented monitoring begins with an interpretable binary early-warning layer before moving toward richer multimodal fusion and domain-specific military validation. The paper therefore contributes a de-fensible first methodological step toward future artillery-related readiness support based on multimodal wearable analytics.

Paper Nr:	37
Title:	Robust Policy Evaluation and Adaptive Deployment in Cooperative Multi-Agent Reinforcement Learning under Non-Stationary Noise
Authors:	Lucas G. M. Castro and Agustin A. O. Diaz
Abstract:	Cooperative multi-agent reinforcement learning (MARL) is often validated under fixed or clean conditions, although deployed agents may face noisy and non-stationary operating regimes. This mismatch raises a practical question: when trained policies exhibit different tradeoffs between nominal performance and robustness, how should they be evaluated, selected, and deployed under changing noise conditions? This paper presents a lightweight evaluation protocol for robust policy assessment and deployment under non-stationary noise. Candidate policies are evaluated under both nominal and perturbed conditions, and their behavior is summarized through compact metrics capturing aggregate performance, degradation under noise, and sensitivity to temporal reordering. The analysis reveals distinct operational regimes: in our controlled benchmark, the best aggregate policy achieves S = −88.85, whereas the policy with the smallest degradation achieves G = 0.99, showing that different robustness criteria can lead to different policy choices. We further compare static, adaptive, and mixing-based deployment strategies under non-stationary noise schedules. Fixed mixing achieves the strongest forward-schedule return, improving over the nominal-best static policy from −1615.22 to −1236.87. However, under schedule reversal, soft adaptation yields the lowest temporal asymmetry (58.62, normalized asymmetry 0.038), while hard adaptation remains more path-sensitive (509.10). These findings suggest that robustness in cooperative MARL should be understood not only as a property of individual policies, but also as a property of how policies are evaluated, selected, and combined under changing noise conditions.

Paper Nr:	38
Title:	Machine Learning Utilization for Multimodal Respiration Tracking and Endotracheal Tube Monitoring
Authors:	Anya Trutschl and Urska Cvek
Abstract:	Endotracheal intubation is widely used in intensive care units for unconscious and semi-conscious patients who cannot breathe independently. Unplanned extubation and tube migration of mechanically ventilated patients are a common occurrence which result in longer ICU stays, unnecessary deaths, and excess avoidable health costs. Current endotracheal tube placement tracking is achieved through continuous capnography (CO2 tracking) and pulse oximetry, both of which suffer from false alarms up to 70% of the time, causing healthcare worker alarm fatigue. Capnograms are sensitive and need frequent replacements, so a simpler, redundant, cost-effective and more efficient system is needed. Such system would have to automatically monitor tube position, track the patient’s breathing, and use the breathing rhythm to identify correct tube position. This research project utilizes simple, readily available sensors and a microcontroller, and through testing shows that endotracheal tube humidity, temperature, and pressure, combined with breathing sounds, supports redundant multimodal detection of respiratory rate and breathing rhythm. After obtaining data from human subjects tests, the results show that the detection modes (humidity, temperature, and pressure) provide increased accuracy and simplification of respiratory monitoring. Spectral features extracted from breathing sounds alone nearly perfectly classify inhales and exhales using a machine learning algorithm, providing a novel endotracheal tube monitoring technique.