ABSTRACT. Tackling grand scientific challenges in an industrial research environment requires a unique convergence of long-term vision, massive compute infrastructure, and multidisciplinary collaboration. In this talk, we explore how scientific research is conducted at Google DeepMind and examine the emerging paradigm shift toward Agentic Science. We introduce the open-source Science Skills suite, an agentic framework that equips language models with quality-verified tools across genomics, proteomics, and scientific literature. Through real-world biological case studies, we demonstrate how structured skills allow efficient models to match or exceed the reliability of unassisted frontier models while achieving substantial token savings. Finally, we discuss open challenges in tool maintenance, benchmark design, and agent steerability.

ABSTRACT. The research examines artificial intelligence tools as instruments for facilitating English language acquisition within profession-specific simulated communicative environments. The research addresses a professionally competent adults who possess substantial work experience and domain expertise yet face the urgent communicative demand of operating in English-medium professional contexts — whether as a consequence of international relocation, employment with a foreign company, or integration into a globally oriented workplace. The study argues that conventional language learning methodologies and generic communicative situations frequently prove ineffective for this learner cohort, as they impose an additional cognitive burden of adapting to unfamiliar contexts and lexical domains unrelated to the learner's professional background. The proposed approach centers on the construction of profession-specific simulated language environments that mirror the authentic communicative situations encountered in the learner's field of activity. Such environments are posited to reduce affective barriers and emotional stress, enable focused engagement with target grammar structures, sentence construction patterns, and oral production skills, and eliminate the distraction of processing contextually irrelevant linguistic input.

ABSTRACT. This paper explores the socio-economic preconditions for the development of online dispute resolution (ODR). The study aims to identify typical coun-try profiles based on socio-economic indicators and to interpret ODR levels within the detected clusters. The empirical basis includes Eurostat data for 30 countries described by Population, HICP, Unemployment, GNI, GDP, Tax, and ODR. The study applies data preprocessing, feature standardization, and comparative clustering using KMeans, Agglomerative Clustering, and Gauss-ian Mixture. The best internal validation metrics were obtained for the two-cluster configuration (Silhouette = 0.6304, Davies–Bouldin = 0.2454), while a three-cluster solution was selected for deeper interpretation. The agreement between KMeans and Agglomerative Clustering reached 0.869 by Adjusted Rand Index, and the first two principal components explained 71.7% of the variance. The results revealed three country profiles with mean ODR values of 2556.33, 29735, and 65, respectively. The proposed approach can be used for comparative assessment of the environment for ODR development.

ABSTRACT. We describe the Anthill language, a many-sorted algebraic specification language in the OBJ and Maude tradition. It uses namespaces, sorts, and operations as its principal means of source organization, and λProlog-inspired Horn-clause rules with higher-order pattern fragment unification. Anthill is designed for gradual formalization — the idea, analogous to gradual typing, that a codebase can be accompanied by expressions at different phases of formalization, from prose description to machine-verified proof, without rewriting them between phases. Declarations become facts in a knowledge base, alongside any free-text description blocks attached to them; informal intent and formal content coexist in the same file and are both queryable by agents and tools. A fact of a given sort may be backed by an external store such as a database or file stream. Anthill specifications can be realized in two ways. The first maps them to a host language; each Anthill sort and operation is paired with a host-side definition. The second runs specifications directly in the Anthill kernel through an expression evaluator. Side effects are declared as sorts alongside data; at runtime each declared effect is interpreted by a handler bound to a concrete implementation. Proofs are attached to specification rules and support multiple backends — internal resolution, external solvers, randomized testing — with the chosen backend determining the trust level recorded for the claim.

ABSTRACT. Passing is one of the key actions for building tactics in football and one of the most common events on the pitch, which allows us to use a limited amount of spatiotemporal data for analysis. Our study aims to understand who is the most likely receiver of a pass initiated at a given moment, as well as what the chances are of the pass being successful based on player movement dynamics on the pitch. We incorporate game state and contextual information, such as possession and phase of play, into the modeling process. We use spatiotemporal data to construct custom features and compare the performance of both classical machine learning methods and latent state-space probabilistic models with Bayesian inference. Predictions and evaluation produced by this method outperform the proposed baseline and provide reasonable assumptions and explainability for a given game context using empirical data from 7 professional football matches. The results demonstrate the predictive power of the proposed features and can be used as a foundation for future modeling.

ABSTRACT. In today’s business environment, the exponential growth in data volumes demands advanced visual analytics tools to support managerial decision-making. This article explores the capabilities of large language models (LLMs) in interpreting graphical financial information. The relevance of this work stems from the need to overcome technological barriers to analytical accuracy by incorporating unstructured data, such as infographics and charts, which are becoming a universal language of communication in business analytics. The study aims to compare the effectiveness of the multimodal models ChatGPT-4.1 and Claude Enterprise Standard in recognizing and converting visual data into structured numerical indicators. The methodology is based on developing a system architecture for processing multimodal queries, including data validation and transmission via cloud storage or the Base64 format. The experimental part covers the analysis of 100 graphical objects (line and pie charts), with results evaluated using MAE, RMSE, and Bias metrics. The scientific novelty lies in applying new approaches to intelligent data analysis while empirically substantiating the advantages and limitations of modern LLMs when working with financial visualizations of varying complexity, particularly those without numerical labels and under conditions of incorrect input data. The main results demonstrate a significant advantage of ChatGPT-4.1: for linear graphs, MAE was 0.13 versus 1.27 for Claude. It was found that ChatGPT better interprets curve shapes and coordinate grids, while Claude more often makes errors in identifying categories. The conclusions state that integrating NLP models into business analytics systems enables effective automation of report processing but requires mandatory verification of results to avoid systematic biases.

ABSTRACT. The evaluation of academic staff activity remains a challenging task for higher education institutions because current assessment procedures are often fragmented, labor-intensive, and dependent on manually verified data. In practice, this leads to inconsistent scoring, limited transparency, and delayed managerial decisions. This paper addresses the problem of building a reliable and repeatable rating assessment mechanism for academic staff within a virtual educational environment. The proposed solution combines academic, methodological, organizational, and research-oriented indicators and supports the use of scientometric data to improve the objectivity of assessment. The developed module enables automated collection, verification, aggregation, and reporting of rating data, which reduces manual workload and improves the comparability of results across staff members and structural units. The obtained results demonstrate that the proposed approach increases the transparency and consistency of academic performance evaluation, while reducing manual data processing efforts. In addition, the system enables multi-level analytical assessment, supporting decision-making at the departmental, faculty, and institutional levels. This contributes to improved governance of academic processes and provides a scalable framework for data-driven management in higher education institutions.

ABSTRACT. The scale and speed of online disinformation increasingly exceed the capacity of manual verification and fact-checking. This challenge is especially acute in active conflict settings, where disinformation can shape public opinion, undermine institutional trust, and influence humanitarian outcomes. This paper investigates whether semantic similarity links can improve graph-based detection of pro-Russian disinformation narratives in the Ukrainian-Russian Telegram ecosystem.

We construct a heterogeneous directed graph of 81,369 Telegram messages from 95 public channels, representing channel-message provenance, explicit forwarding relations, and semantic similarity edges derived from multilingual BAAI/bge-m3 embeddings. Message-level narrative labels are obtained through weak supervision using few-shot LLM prompting against the TeleNarratives dataset taxonomy. We train a two-layer HeteroGraphSAGE model with LSTM aggregators under four edge configurations and compare it with a RoBERTa text-only baseline using a channel-aware chronological split.

Results show that semantic similarity edges provide an additional signal for narrative detection beyond explicit forwarding structure alone. The best full-graph configuration, using channel-posting and semantic similarity edges, achieves F1 = 0.82, MCC = 0.60, and PR-AUC = 0.81, closely matching the RoBERTa baseline. A restricted subgraph containing only messages involved in semantic similarity relations reaches F1 = 0.89 and PR-AUC = 0.94, suggesting that semantic-neighborhood structure is highly informative for identifying narrative-bearing messages. These results position semantic graph construction as a useful complement to forwarding-based diffusion analysis for monitoring disinformation narratives on Telegram.

ABSTRACT. In the paper, the task of detecting potentially coordinated activity in com-ments of Telegram channels is considered under the conditions of the ab-sence of an explicit reply structure (reply_to). An explainable two-stage tri-age approach is proposed: at Stage A, a rule-based module forms an inter-pretable risk scoring of dyad interactions (a proxy “comment→comment”) based on time lags, burst features, the number of unique authors, and textual similarity; at Stage B, a confirmation model performs classification/re-ranking of candidates within an enriched candidate pool with control of fea-ture leakage and class imbalance. Experiments were conducted on a corpus of 247,662 comments from 3,390 threads and 8 channels, where 2,041,964 dyad pairs were generated. Stage A converts a rare weak-positive event (0.421% of all dyads) into a manageable Top-8000 pool with a signal con-centration of 57.675%. On the candidate pool test, the selected Bagging_DT model achieved ROC-AUC=0.7004, PR-AUC=0.7215, and at the optimal threshold F1=0.7399 (Precision=0.6265, Recall=0.9034); in triage mode, Precision@50=0.78 and Precision@100=0.81 were obtained, which signifi-cantly exceeds the Stage A heuristic as a ranker. Additionally, explainability is shown through the analysis of rule firings and practical suitability by the p95 latency of batch inference. The results confirm the expediency of the two-stage scheme “explainable selection → confirmation/re-ranking” for monitoring potential information operations in Telegram discussions.

ABSTRACT. In this paper, we propose a compositional two-layer stochastic model as a formal basis for simulation-oriented analysis of collective decision-making in networked communities. The model separates the latent opinion dynamics from a non-invasive observation layer. Four primitives are composed into a single population-level Markov kernel on the agent state space: activation, pairwise dialogue, exposure aggregation, and internalization. We then characterize absorbing states coordinate-wise and isolate two structural sufficient conditions, namely inactivity and frozen dialogue. At any latent absorbing state, deterministic voting yields a constant observed outcome, while stochastic voting yields i.i.d. outcomes with an explicit push-forward law. This separation distinguishes the role of the dynamics from that of the decision procedure. As a worked example, an extended pairwise dialogue mechanism with persuasion–resistance attributes is embedded into the framework; its nopersuasiveness regime (with strictly positive resistance) provides a concrete deadlock case.

ABSTRACT. The real estate rental market is a highly dynamic environment influenced by diverse factors, including physical property characteristics, infrastructure accessibility, and socio-economic sentiments reflected in textual descriptions. While advanced machine learning algorithms, particularly ensemble methods like XGBoost and Random Forest, have demonstrated superior predictive accuracy compared to traditional regression models, their "black-box" nature remains a significant barrier to adoption in decision-support systems for landlords, tenants, and city planners. This study proposes a comprehensive framework for housing price prediction that integrates high-performance gradient boosting models with Explainable Artificial Intelligence (XAI) techniques. Using a dataset of rental advertisements, we pre-processed multi-source data from Kyiv rental market, including numerical features and location-based metadata. To address the interpretability challenge, we implemented SHAP (Shapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations). SHAP was utilized to provide a global perspective on feature importance, revealing how factors such as square footage and location-specific amenities consistently influence price trends across the entire market. Conversely, LIME was applied to provide local explanations, offering transparency for individual predictions and highlighting why a specific rental unit was valued at a particular price point. The results indicate that while CatBoost achieves the highest accuracy (lowest MAE and RMSE), integrating XAI tools significantly enhances the model's transparency. This dual approach not only improves technical performance but also fosters user trust by providing actionable insights into the underlying drivers of real estate value. The proposed framework serves as a robust ICT solution for real-world industrial applications in the property technology sector.

ABSTRACT. Technology has the potential to allow patients to reintegrate into society after traumatic events leading to amputations. In addition to the functional challenges associated with amputation, chronic neuropathic pain can further hinder these patients’ quality of life. Prof. Ortiz Catalán will lecture on novel but clinically viable technologies to restore quality of life to patients with limb loss. Using direct skeletal attachment via osseointegration, neuromuscular interfaces, and machine learning, his work resulted in the first artificial arm integrated directly into the patient’s bone, nerves, and muscles. Patients implanted with this system are also provided with intuitive sensory feedback via direct nerve stimulation. Direct skeletal attachment via osseointegration, along with control and sensory feedback via implanted neuromuscular electrodes, can now be provided in a self-contained prosthesis for use in daily life. Prof. Ortiz Catalan will also discuss how motor decoding technology, in combination with virtual reality, can be used to treat Phantom Limb Pain, along with a novel theoretical framework for the condition and its treatment. He will also describe how such technologies have been used for the functional and pain rehabilitation of severe sensorimotor impairments.

ABSTRACT. The rapid and pervasive digitalization of higher education administration demands sophisticated, automated tools capable of systematically evaluating multi-dimensional academic performance records. Traditional approaches to tracking and analyzing faculty research contributions, instructional output, and professional milestones are fundamentally constrained by high time consumption, siloed transactional data structures, and an over-reliance on static, purely quantitative numeric aggregates. This paper introduces a comprehensive full-scale architectural design, end-to-end algorithmic implementation, and extensive empirical validation of an intelligent, AI-driven academic analytics module natively integrated within the university Virtual Learning Environment (VLE) platform "KSU24". The developed module orchestrates the automated aggregation, sanitization, and qualitative synthesis of highly heterogeneous faculty performance vectors organized across 14 distinct activity domains. By implementing a decoupled backend infrastructure integrated with leading state-of-the-art commercial and open-weights Large Language Models (LLMs)—specifically GPT-4o, Gemini 1.5 Pro, and Claude 3.5 Sonnet—via a strict deterministic software pipeline, the platform transforms multi-layered relational database entities into highly coherent, actionable, narrative analytical profiles under a robust semi-automated Human-in-the-loop governance matrix. To suppress the inherent non-determinism and systematic semantic hallucinations characteristic of foundational generative APIs, we propose a specialized prompt engineering strategy leveraging role-restricted boundaries, isolated context windows bounded by customized XML tags, and schema-driven data ingestion rules. Long-term field testing conducted over 346 authentic faculty records at Kherson State University confirms absolute data consistency (0.00% factual error rate), complete structural alignment, and significant operational velocity optimization, compressing the typical manual report composition lifecycle from 40–60 minutes down to 2–5 minutes per profile. Furthermore, we outline a mathematical formalization of the semantic translation layer and propose a predictive extension framework designed to simulate long-range career trajectory modeling for institutional talent development and strategic capacity forecasting.

ABSTRACT. Wi-Fi sensing is a promising way to turn everyday wireless infrastructure into a practical sensing tool, but progress in this area depends on results that can be reused and compared fairly. Single-link Wi-Fi sensing results are difficult to reuse because later comparisons remain valid only when the baseline setup, experiment selection criteria, and evaluation rules are fixed in advance. This paper proposes a new method for constructing a reusable baseline experiment set for a single-link Wi-Fi sensing setup built around an ESP32-C5 microcontroller node. The method defines the physical and radio conditions, executes a predefined experiment program, retains experiments only by explicit inclusion criteria, separates accepted and incomplete experiments, interprets control data conservatively, and keeps the block-wise evaluation split unchanged in later experiments. This combination brings setup definition, experiment selection, control interpretation, and later reuse under the same setup into one baseline-construction method. Within the selected four-study comparison set used here, the method states all five method elements explicitly and reaches 4.0x the mean comparator completeness and 2.5x the strongest comparator completeness. Applied to the same-day campaign, the method screened 10 attempted experiments and retained one baseline experiment, one control experiment, and one occupancy experiment. The method then enables a controlled comparison in which compact features reduce the representation size from 628.25 to 40.00 bytes per decision window (6.37% of the packet-summary representation size; 15.71x smaller) on the accepted occupancy experiment, while binary balanced accuracy decreased by 0.0112. The contribution is therefore not a cooperative sensing result, but a documented same-day single-link baseline-construction method and a fixed basis for later within-setup comparisons. This paper is useful as a reproducible basis for later studies on controller-managed cooperative measurement coordination, local feature extraction with central fusion, and stability-aware configuration maintenance in multi-access-point Wi-Fi sensing systems.

ABSTRACT. The rapid advancement of artificial intelligence (AI) technologies, particularly large language models (LLMs) and rule-based conversational agents, has opened new possibilities for augmenting counseling practices within social work. However, there remains a significant gap between the technological capabilities of AI-driven tools and their systematic integration into social service delivery. This paper proposes a conceptual pipeline for embedding AI-assisted counseling into social work practice and presents preliminary findings from a mixed-methods pilot study. The theoretical component draws on the Augmented Social Worker framework, human-in-the-loop paradigms, and established counseling models — including cognitive-behavioral and solution-focused approaches — to construct a four-stage pipeline: (1) automated intake and needs screening, (2) AI-facilitated psychoeducational dialogue, (3) risk detection and escalation to a human professional, and (4) post-session analytics and case documentation support. The pilot study (N = 42) employed a convergent mixed-methods design combining a structured questionnaire measuring perceived usefulness, trust, and ethical concerns with semi-structured interviews exploring practitioner attitudes toward AI integration. Preliminary quantitative results indicate moderate-to-high perceived usefulness scores (M = 3.7/5.0) alongside persistent concerns regarding data privacy and algorithmic transparency. Qualitative analysis revealed three dominant themes: cautious optimism toward routine task automation, resistance to AI involvement in crisis intervention, and demand for culturally and linguistically adapted tools. The findings suggest that AI-assisted counseling holds practical promise for reducing workload and improving service accessibility, provided that ethical safeguards, professional oversight, and contextual adaptation are embedded into the system design. The proposed pipeline offers a structured reference model for researchers and practitioners seeking to operationalize AI integration in social work settings.

ABSTRACT. Abstract. The article examines the role of open science as a mode of organizing research activity that ensures the rapid, transparent, and ethical exchange of research data, fosters high-quality scholarly communication and collaboration between academic and non-academic environments, and supports the storage, sharing, processing, reproducibility, and reuse of digital research objects. The current state of regulation governing the implementation of open science practices is characterized; in particular, the regulatory and legal framework for the development of open science in Ukraine is outlined, and national as well as institutional policies for integrating open science practices into the educational and research activities of higher education institutions and research organizations are considered. Foreign and domestic scholarly sources on open science are analyzed. The concept of higher education lecturers’ readiness to implement open science practices in research activity is substantiated and defined; the structure of this readiness is elaborated as a unity of three components (motivational, informational, and operational), and levels of its development (high, medium, and low) are identified. The study analyzes, synthesizes, and presents the findings on the levels of readiness among the third-cycle (educational and scientific) higher education students, academic staff, and researchers at Borys Grinchenko Kyiv Metropolitan University to implement open science practices in their research activities.

ABSTRACT. This paper addresses the problem of feature vector extraction for solar panel defect analysis using a modified MobileNetV2 architecture. Instead of using MobileNetV2 as a final classifier, its standard classification head is replaced with an embedding head that generates a 256-dimensional feature vector as the main model output. An auxiliary classification branch trained with CrossEntropyLoss provides the supervised learning signal and is used only for proxy evaluation of the discriminative information contained in the learned representation.

The experimental study was conducted on a combined dataset constructed from three publicly available solar panel image datasets and containing six visual conditions: Bird-drop, Clean, Dusty, Electrical-damage, Physical-damage, and Snow-covered. In addition to auxiliary classification metrics, the learned embeddings were evaluated directly using cosine-distance analysis, the Silhouette Score, and nearest-neighbor retrieval metrics. The cosine-based Silhouette Score reached 0.8534, while the mean inter-class cosine distance was 10.27 times greater than the mean intra-class distance. Precision@1 reached 0.9841 and mAP@5 reached 0.9771. The auxiliary classification branch achieved an Accuracy of 0.9781 and a macro F1-score of 0.9794.

The obtained results demonstrate that the modified MobileNetV2 generates compact feature vectors with a well-separated and class-consistent structure, which makes them suitable for subsequent embedding-based analysis, including similar-defect retrieval and other downstream processing tasks.

ABSTRACT. Artificial intelligence (AI) adoption has reached unprecedented scale. 78% of organizations now report using AI in at least one business function, up from 55% a year earlier. Meanwhile, generative AI has reached 53% of the global population within just three years of its public release, a faster diffusion than the personal computer or the internet [1, 2]. However, adoption remains uneven across sectors and mediated by human factors (trust, anxiety, perceived usefulness, resistance) that classical measurement models do not capture. The paper proposes an integrated AI adoption framework with seven new cross-cutting metrics — AUAR, FUD, TTC, AER, AV, API, and ROAI — each analytically decomposed into sector-level adoption constructs drawn from recent peer-reviewed literature, producing a literature-grounded dependency mapping. The framework integrates classical technology acceptance theory [4, 5], macro-level diffusion evidence [1, 2], sectoral adoption evidence [3], human-factors literature [10, 11, 12], and the legacy of intelligent systems quality evaluation and defect-profile life-cycle approaches [16, 17]. The framework is further parameterized by an adoption-intent mode (augmentation versus replacement) with usage-economics measurement, and a three-stage empirical validation plan is outlined.

ABSTRACT. The transformational potential of Artificial Intelligence (AI) presents both un-precedented risks and unique opportunities for the global cybersecurity domain. As threat actors integrate generative AI into their offensive toolkits to automate polymorphic malware and execute sophisticated social engineering campaigns, traditional deep learning defense mechanisms exhibit fundamental limitations in explainability and conceptual grounding. This overview systematically investi-gates the contemporary cyber threat landscape spanning 2024–2025, detailing the evolution of Data-Oriented Attacks (DOA) and AI-enhanced social engineering. To address these vulnerabilities, we propose the Neuro-Symbolic (NeSy) ap-proach, which synergistically combines the statistical pattern-matching capabili-ties of neural networks with the rigorous logic of symbolic reasoning. Through formal semantic analysis, benchmarking of state-of-the-art Large Language Mod-els within Security Operations Centers (SOCs), and the application of the Grounding-Instructibility-Alignment (G-I-A) framework, this paper evaluates the efficacy of NeSy architectures. Finally, we outline a comprehensive research roadmap to 2030 directed toward the realization of fully cognitive, self-healing cyber defense systems.