ABSTRACT. This is the opening talk given by the Program Chairs. It informs about conference statistics and program.

ABSTRACT. What do we actually know about the world we inhabit? What do we think we know, but actually don’t. This seminar uses AI-based super-resolution in remote sensing and Earth Observation as a practical entry point into a deeper question: what makes scientific knowledge possible, and where do its limits really lie? We will begin with modern AI systems that infer high-resolution structure from low-resolution satellite observations, showing how technological progress often depends on constraints, priors, and hidden assumptions rather than unlimited information. From there, we will discuss the foundations of mathematics and science itself: the assumptions behind number, infinity, physical law, measurement, objectivity, and the role of observer’s frame of reference. Drawing on recent insights on relational finitude, the illusion of absolutes, and the metaphysical commitments hidden inside formal definitions, we will examine how scientific intuition is shaped by cognitive horizons, historical ego-centrism, and inherited abstractions. The central message is that technology advances precisely when we learn to question the boundaries that earlier frameworks treated as fixed. Super-resolution is therefore not only an imaging technique, but a case study in how science sees beyond its own inherited resolution.

ABSTRACT. Educational institutions are adopting generative AI faster than the governance frameworks required to ensure academic integrity and equitable access. Existing policies often fail to translate industry-grade AI standards into operational classroom controls, leaving a gap between abstract ethical guidelines and daily practice. This paper addresses this gap by proposing a lightweight socio-technical governance framework derived from international standards (ISO/IEC 24028:2020, ISO/IEC 25059:2023, and ISO/IEC 42001:2023). Using a structured four-phase methodology, we mapped core AI trustworthiness characteristics, such as transparency, accountability, and fairness to specific academic requirements. The resulting framework was formally adopted as the official AI Usage Policy at Yuriy Fedkovych Chernivtsi National University (August 2025). A mixed-methods preliminary evaluation—qualitative pilot feedback and a university-wide staff survey (N=106)—indicates the feasibility of disclosure-based governance: respondents broadly endorsed disclosure requirements and instructor discretion, yet 76.4% remain concerned about student over-reliance on generative AI, underscoring the need to couple disclosure with assessment-design and AI-literacy work. The findings provide a standards-based, replicable starting point for universities seeking to operationalise trustworthy AI in regulated, low-resource settings.

ABSTRACT. The digital transformation of higher education in Ukraine, exacerbated by the challenges of the pandemic and martial law, requires a rethinking of approaches to the professional development of academic staff. The article examines the impact of the digital learning environment on the formation of digital competencies of teachers of a higher education institution, taking into account the role of digital thinking and digital adaptation skills as mediators. The aim of the study is to study the structural relationships between the digital learning environment (DLE), digital thinking (DM), digital adaptation skills (DAS) and teaching digital competencies (TDC). The methodology is based on structural equation modeling using the partial least squares (PLS-SEM) method based on the results of a survey of 167 teachers of Kherson State University. The scientific novelty lies in the development of a multidimensional model that proves that the digital environment affects teaching digital competencies not directly, but through the mediator variables of digital thinking and digital adaptation skills. The results demonstrate that the proposed model explains 75.5% of the variation in teaching digital competencies. Digital adaptation skills were found to be a stronger mediator (coefficient 0.568) than digital thinking (0.301). The hypothesis of a direct impact of infrastructure (DLE) on the level of competencies was rejected, which emphasizes the critical role of the teacher’s internal readiness for change. A three-level adaptation model (individual, professional and system levels) is substantiated and practical cases of implementing policies on the use of AI and continuous development programs for academic staff as effective ICT solutions in the design of digital teaching competencies are presented.

ABSTRACT. The rapid expansion of artificial intelligence in scientific writing, analysis, knowledge production and technology transfer (postpublication process) has intensified debate over how to preserve research integrity in AI-assisted scholarship. Existing attribution taxonomies improve transparency by identifying where AI has been used, but they do not sufficiently capture the extent to which AI participates in specific research tasks. This fact creates a significant conceptual and ethical gap: disclosure alone cannot determine whether human authors retained adequate intellectual control over the research process. The purpose of this article is to demonstrate that attribution of AI is a necessary but insufficient condition for maintaining research integrity and to propose a mathematical model that evaluates AI involvement not only by attribution, but also by the role AI plays within each task. The study uses a conceptual-analytical design grounded in formal mathematical modeling. A role-sensitive integrity model is developed in which scientific activity is divided into a set of research tasks, each assigned an epistemic weight, a level of attribution completeness, a degree of human verification, and a graded coefficient of AI role intensity derived from an assessment scale. The model compares an attribution-only approach with a role-sensitive integrity framework to test whether disclosure alone is a sufficient condition for research integrity. The model demonstrates that attribution- only logic is analytically limited because it treats all disclosed AI uses as equivalent, regardless of whether AI performed minor linguistic assistance or substantial intellectual work in data interpretation, argument construction, or conclusion drafting. By contrast, the proposed role-sensitive integrity index shows that integrity depends on the interaction of four variables: attribution completeness, AI role intensity, human verification, and the epistemic significance of each task. The formal analysis shows that even under full AI attribution, research integrity declines whenever AI plays a substantial role in high-weight research tasks without corresponding human oversight. Thus, attribution functions as a transparency condition, but not as a sufficient safeguard of integrity. The findings suggest that current approaches to AI disclosure in research are too narrow if they rely only on taxonomic attribution. Preserving research integrity requires a more differentiated framework that evaluates not merely the presence of AI, but the depth of its participation in each component of scientific work. The proposed model contributes to the field by offering a formal basis for moving from disclosure- based governance to role-sensitive assessment of AI use. More broadly, it implies that future standards for academic integrity, editorial policy, and research ethics should incorporate structured evaluation of AI agency alongside attribution, thereby strengthening human accountability in AI-assisted scientific activity.

ABSTRACT. Narrowband IoT (NB-IoT) over non-terrestrial networks (NTN) extends cellular reach to maritime, agricultural, and critical-infrastructure assets that were previously out of coverage, but introduces a security gap: resource-limited IoT terminals cannot host traditional intrusion de- tection systems. Classication under these constraints requires an alter- native compatible with the memory and latency limits of the deployed module. This paper presents a quantized intrusion detection system that ts this constraint. The Quantized JohnsonLindenstrauss (QJL) rota- tion of the TurboQuant family is adapted from its large-model setting to the oine compression of a small multilayer perceptron and paired with a per-row LloydMax codebook. We release an NB-IoT NTN trac dataset generated with our ns-3 fork under a 24-hour LEO-pass sched- ule. Over 30 independent training and rotation seeds, the 3-bit detector reaches weighted F1 = 0.9987±0.0017 at a deployed footprint of 1.88 kB.

ABSTRACT. Smart contracts are critical components of modern blockchain platforms, but traditional post-hoc verification methods, such as testing and manual audit-ing, are often insufficient to prevent critical errors in complex distributed sys-tems. This paper presents a correct-by-construction approach to smart con-tract development based on formal modeling within the Insertion Modeling System (IMS). Contract behavior is formalized using algebraic specifications that define agents, their actions, and interactions. The model is subsequently analyzed through simulation and trace exploration to confirm the reachability of correct final states prior to implementation. Building on the validated model, a multi-stage, semantics-driven transformation pipeline is introduced to generate executable Solidity code from IMS specifications. The approach enables early detection of logical errors during the modeling phase and pro-vides correctness guarantees at the model level, thereby improving the relia-bility and security of smart contract development. Its effectiveness is demon-strated through the formal modeling and code synthesis of an English Auc-tion smart contract, confirming the applicability of the proposed approach.

ABSTRACT. High-throughput lossless compression is a critical enabler for Earth observation microsatellites operating at multi-Gbit/s scanner data rates. This paper proposes and benchmarks three spatial partitioning strategies – Horizontal, Vertical, and Block-based – for a multi-core JPEG-LS encoder targeting the Artix-7 FPGA and introduces a new method for selecting the required number of encoder cores and the operating clock frequency from first principles. The method derives both parameters directly from the scanner pixel rate, making core-count selection independent of pixel bit depth. Because a one-pixel-per-clock LOCO-I core closes timing at about 53 MHz on the Artix-7, the method sizes the design at the device-achievable clock rather than an idealized one: for the Sich-3O pixel rate P = 518 Mpix/s the rule N = P / F yields twelve lanes at 45 MHz, a configuration that was synthesized and timing-closed (worst negative slack +0.65 ns) at 540 Mpix/s ≈ 4.32 Gbit/s, meeting the 4.15 Gbit/s requirement; the same design reaches 853 Mpix/s ≈ 6.83 Gbit/s at the 53.3 MHz device fmax. Core count is independent of pixel bit depth, so this sizing holds for 8-, 12-, and 16-bit pay-loads. All results are taken from post-route implementation reports and compared with commercial and academic implementations.

ABSTRACT. Large language models (LLMs) achieve strong performance on standard benchmarks but may show unstable behavior under small meaning-preserving input changes. This issue is particularly challenging for Ukrainian, a morphologically rich and relatively low-resource language, where synonym substitution must preserve word sense, inflection, and fluency. We evaluate the robustness of LLMs to Ukrainian synonym-substitution attacks using transferred adversarial examples generated against fine-tuned encoder models. Experiments on three classification tasks---sentiment analysis, news classification, and manipulation detection---show that such perturbations frequently transfer to LLMs, especially for sentiment and manipulation tasks. However, manual analysis of 300 adversarial failures reveals that standard robustness metrics substantially overestimate true brittleness: only 19\% of failures correspond to strict meaning-preserving prediction flips, while many are explained by semantic drift, grammatical errors, or label ambiguity. To improve perturbation quality, we introduce a WSD-based filtering method that increases the share of semantically meaningful perturbations. Our results suggest that the evaluated LLMs are sensitive to Ukrainian lexical variation. Overall, we provide a robustness evaluation focused on Ukrainian that highlights how lexical variation, morphology, and perturbation validity affect LLM behavior under synonym substitution.

ABSTRACT. Handwritten text generation (HTG) conditioned on writer style has been widely studied for Latin scripts, but remains underex- plored for low-resource and non-Latin writing systems, leaving open how well existing models generalize beyond the Latin domain. Cyrillic— particularly Ukrainian—lacks both large-scale writer-labeled datasets and empirical evidence of such generalization. To address this gap, we construct a Ukrainian handwritten word-image dataset with over 120,000 samples from over 300 writers using connected-component segmenta- tion, quality filtering, and targeted oversampling of underrepresented Ukrainian characters. We retrain DiffusionPen—a MobileNetV2 triplet- loss style encoder with a CANINE-conditioned latent diffusion U-Net— on this dataset without architectural modification, testing direct trans- fer from Latin to Cyrillic. We evaluate cross-domain style transfer in three settings: cross-lingual transfer from IAM English samples, zero-shot transfer to an early 20th-century Ukrainian manuscript, and few-shot imitation of contemporary writers. The model produces legible, style- consistent word images, indicating that few-shot latent diffusion mod- els generalize beyond the Latin-script domain. We release the dataset, trained models, and evaluation protocol as a reproducible benchmark for writer-aware Cyrillic HTG, providing a foundation for extending stylized HTG to other underrepresented writing systems.

ABSTRACT. Vision--language models hallucinate, and explanations for why they do so range from language-prior dominance to spurious visual cues learned during training. We probe these mechanisms through input-level image perturbation, applied before inference and requiring no modification to the model itself. Across four models (Qwen2-VL, LLaVA, PaliGemma, IDEFICS2) on POPE and CCEval, we find that response to perturbation partitions the models into two distinct hallucination profiles: PaliGemma and LLaVA hallucinations are predominantly fragile to high-frequency removal (Gaussian blur), while Qwen2-VL and IDEFICS2 hallucinations are predominantly fragile to discrete pixel-level corruption (salt-and-pepper, dropout). This pattern is consistent across POPE difficulty splits, suggesting an architectural rather than benchmark-specific property. Building on this analysis, we introduce a targeted patch-level perturbation guided by GradCAM and cross-attention, which corrects up to 24\% of object hallucinations while flipping fewer than 2\% of correct answers, a favourable benefit/harm ratio that distinguishes it from full-image perturbation, where the same correction comes at 8--15\% harm. Finally, we show that the discriminative-mode profile does not transfer to open-ended captioning, where output-length effects confound count-based metrics and PaliGemma's perturbation response in fact reverses direction.

ABSTRACT. Ensuring the cybersecurity of web-oriented systems is one of the key chal-lenges in modern software engineering. In practice, the selection of checks and tools is largely determined by the qualifications of a particular special-ist rather than the characteristics of the system under study, which affects both the completeness of vulnerability coverage and the reproducibility of results. This paper presents a penetration testing model for web-oriented systems based on the analysis of project documentation and the application of artificial intelligence methods for test plan generation. Structurally, the model comprises four sequential blocks whose interrelations are defined through sets of system parameters, vulnerability classes, test scenarios, and tools. A concrete example demonstrates that the model enables automated generation of a targeted set of checks and tools for a given technology stack. The application of the model contributes to improved vulnerability detection completeness, reduced preparation time, and decreased depend-ence of testing outcomes on individual specialist expertise.

ABSTRACT. Abstract. Strategy families in the Iterated Prisoner’s Dilemma (IPD) are usually hand-crafted by domain experts: titles such as tit-for-tat-like, grudger, zero-determinant are attached on the basis of an author’s de- scription of the algorithm rather than its actually-observed behaviour. This makes the resulting taxonomy hard to compare across implemen- tations and to extend to new, learned agents. In this study, expert tags are replaced with a purely data-driven family construction. Each strat- egy is described by a behavioural fingerprint, built either from matches against a fixed panel of ten classical probes (Cooperator, Defector, Tit For Tat, . . . ) or from per-step rewards and cooperation rates obtained when the strategy is played against a panel of independently trained deep reinforcement-learning (RL) agents. We cluster up to 243 strategies from the axelrod catalogue in the classical, RL-based, and combined feature spaces using hierarchical clustering, and study repeatability along five axes: bootstrap re-sampling, choice of the number of clusters K, linkage– metric combination, RL training seed, and feature-space substitution. The classical fingerprint produces a highly stable 4-cluster solution (boot- strap ARI = 0.775, silhouette = 0.572), while the RL fingerprint resolves a finer 11-cluster structure (bootstrap ARI = 0.709) that collects all nine zero-determinant variants in a single exploit-resistant cluster shared with hard defectors (C5 , n=40) – a behavioural niche the classical fingerprint scatters across two unrelated clusters. Under this clustering design (classi- cal K=4, RL-based K=11), the two partitions show strong disagreement (cross-feature ARI = 0.058, NMI = 0.214), consistent with the interpre- tation that probes measure “how does this strategy respond to canonical opponents” while RL agents measure “how exploitable is this strategy by a learned opponent”. Linkage–metric ablation confirms that average link- age with cosine distance is the most internally consistent configuration, and a seed-vs-seed comparison of RL clusters bounds the variability that training stochasticity introduces (ARI = 0.457). Both feature families recover the expert families above chance (ARI = 0.37 classical, 0.48 com- bined on labelled subsets) without ever seeing the labels, supporting the view that behavioural clustering can serve as an objective replacement for hand-curated taxonomies.

ABSTRACT. Constructiveness is harder to annotate than toxicity or stance because it requires context-sensitive ordinal judgments about interper- sonal conduct and substantive contribution. We evaluate whether large language models can act as rubric-based annotators for Ukrainian on- line discussions under a two-axis framework covering Relational Conduct (RC) and Substantive Contribution (SC). Using 300 expert-annotated political discussion items, we test five LLMs in four prompting conditions and compare them against expert agree- ment references and a fine-tuned CORAL XLM-R baseline. Gemini 3.1 Flash Lite performs best: with few-shot prompting and no additional context, it reaches macro QWK 0.584 (95% CI [0.529, 0.633]), above the CORAL baseline of 0.389 and numerically close to the averaged expert pairwise agreement (0.582) [0.537, 0.620], while costing $0.51 per 1,000 items. The same configuration remains weaker on pragmatic discourse- management dimensions (RC2 and SC3), where all five models compress the ordinal scale and miss code-switching cues. Few-shot prompting is usually beneficial and conversational context has mixed effects. Current LLMs are therefore useful but imperfect constructiveness anno- tators: strong enough for aggregate analysis and monitoring, but not reli- able enough for high-stakes individual judgments without human valida- tion. We further show that cross-model disagreement provides a practical uncertainty signal: routing the most uncertain items to human annota- tors improves the reliability of the remaining automated labels (macro QWK 0.612 → 0.672), enabling a controllable human-in-the-loop anno- tation pipeline

ABSTRACT. Foundation models are remarkably capable but fundamentally lack local context, which are the domain-specific, community-specific, and structurally rich knowledge required by real-world applications. In this talk, I will discuss our work on bridging this gap through knowledge graphs, localized foundation models, and domain-aware reasoning. I will show how these approaches enable applications for wildlife conservation research, navigation of Colombia’s Truth Commission archives, and healthcare, the very settings where generic model knowledge falls short and local context is everything.