
Large language model (LLM) agents and agentic artificial intelligence (AI) are entering business process management, yet how process models should frame their behavior remains unclear. This PRISMA 2020 systematic literature review examines how process models, business rules, and standard operating procedures frame LLM-based systems. Searching Scopus, arXiv, AIS eLibrary, and DBLP on 3 October 2026, we screened 1775 records in duplicate and included 126 studies, coded on 18 variables. Process specifications play three roles: context the model reasons over (31%); control that constrains its actions (27%); and governance that allocates work, sequence, and autonomy (42%). The standard notations, Business Process Model and Notation and its decision and case counterparts, appear in 21% of studies, while other formal models, rules, and natural-language procedures dominate, and a third of the systems combine several of them. Each role has a typical form: procedures rendered as text for context, rules enforced by engines for control, and formal models, often with multiple agents and human oversight, for governance. Notation inadequacies are reported mainly for governance, and the evidence is emergent. We contribute a seven-dimension taxonomy (role, secondary role, notation, representation, autonomy, system type, and orchestration), and a research agenda for notations that govern agentic AI.
Exact, deterministic preparation of an arbitrary dense n-qubit state is the data-loading step for quantum signal and image processing, and, without ancillae, a generic target needs at least ⌈(2n − n − 1)/2⌉ controlled-NOT (CNOT) gates. We evaluate the two new Quantum Signal-induced Heap Transform (QsiHT) syntheses, Real and Complex, built on the discrete signal-induced heap transform (DsiHT), inside a nine-configuration benchmark drawn from the uniformly controlled rotation, multiplexor, isometry, Schmidt, and Quantum Shannon Decomposition (QSD) families. Qiskit’s StatePreparation is the only method benchmarked as shipped rather than re-implemented. Every configuration reproduces its Haar-random targets to machine precision, with the Real Synthesis confined to sign-bearing targets. Under noise, one coarse tier is stable, set by the executed two-qubit count; the O(2n) frontier stays well below QSD’s O(4n), and no finer order reproduces across two IBM Heron r2 processors at n = 3. Both syntheses reach the 2n − n − 1 multiplexor-family floor. That count matches StatePreparation and, as built, undercuts every other from-scratch construction from n = 3 up. Under a uniform optimization-level-3 pass Plesch–Brukner falls below it at n = 4 and on real targets at n = 6, and as built, it keeps the lowest constant on complex data among the from-scratch methods above that floor. At large registers, the Real Synthesis has the cheapest build-time cost and the Complex Synthesis handles arbitrary targets.
Identifying employees who may need additional professional-learning support could help organisations allocate training resources, but model comparisons in this area often omit trivial baselines and class-specific metrics. We compare seven classifiers (decision tree, random forest, logistic regression, k-nearest neighbours, XGBoost, LightGBM, CatBoost) with a rule-based and a majority-class baseline under one protocol: stratified hold-out, the same tuning procedure, paired bootstrap confidence intervals, a descriptive repeated cross-validation check, and SHAP and LIME explanations. Experiment 1 predicts a KPI label of undocumented origin for 17,415 employees (35.9% positive); it is an illustrative comparison using the label as distributed, not validated evidence about KPI attainment. Experiment 2 predicts promotion for 54,808 employees (8.5% positive), a related task from the same public source and not an external validation. In Experiment 1, learning models reached macro-F1 of 0.628 to 0.663, against 0.590 for the rule-based baseline; a depth-5 decision tree was 0.005 below the reference random forest (95% bootstrap CI −0.016 to +0.005) but was not shown non-inferior at a 0.01 margin. In Experiment 2, boosting models were best (macro-F1 0.734 to 0.739), the decision tree was clearly worse (−0.030), and only about 35% of promoted employees were identified. Associations with training indicators are not causal, and no effect on decisions was tested.
In education, artificial intelligence (AI) plays an increasingly important role and is contributing to the growth of more adaptive, interactive, and personalized forms of learning. This paper considers this evolution from AI-based education toward cognitive education, from three main dimensions: educational applications, AI technologies, and intelligent educational systems. Based on the review of the literature, the paper first presents the main applications of AI in education such as personalized learning, teaching and instructional support, student monitoring and academic integrity, and experiential and immersive learning. Then, it reviews the main technologies used to support these applications, such as machine learning and deep learning, natural language processing, computer vision, knowledge representation and reasoning, and more recent developments in generative AI. The study also examines how these applications and technologies are implemented in intelligent educational systems and platforms. A cross-dimensional analysis examines the relationships between these three dimensions. Complementing this analysis, we propose six functional orientations based on the conceptual synthesis and examine six cognitive capabilities: perception, learning, context awareness, reasoning, adaptation, and interaction. The proposed framework connects the three dimensions and the six capabilities within a continuous feedback loop. Cognitive Education is thus presented as a conceptual research direction requiring empirical validation. Finally, the article discusses the main technical, pedagogical, ethical, and governance challenges, along with emerging trends and directions for future research. Finally, the article discusses the main technical, pedagogical, ethical, and governance challenges and presents emerging trends and possible directions for future research.
Purpose: This study addresses the small data challenge in additive manufacturing, where destructive tensile testing limits sample size (n = 81), by developing and validating a probabilistic synthetic data generation pipeline. Design, Method, and Approach: A twelve-stage framework couples Gaussian Process Regression (GPR) with three-tier validation. Infill patterns are one-hot encoded, process parameters are sampled uniformly within physical bounds, and synthetic ultimate tensile strength (UTS) is generated with uncertainty-guided noise (γ = 0.3). Fidelity is assessed through distributional, structural, and Train–Synthetic Test–Real (TSTR) tests using XGBoost, Random Forest, and GPR. Results: GPR expanded the dataset 617-fold to 50,000 synthetic observations (1000 used for validation), preserving the dominant infill–UTS dependence, while agreement for other features varied. Models trained on synthetic data reached a physical test R2 of 0.8895 (XGBoost) and 0.9185 (Random Forest), 3.61 and 1.02 percentage points below the experimental baselines (0.9256 and 0.9287); GPR reached 0.9425. Originality: Combining one-hot encoding, bounded sampling, and multi-tier validation, this work shows that probabilistic synthetic data can complement physical specimens when training machine learning surrogates.