<p>Generative Artificial Intelligence has undergone rapid maturation between 2023 and 2025, driven by three converging paradigm shifts: the emergence of multimodal foundation models unifying text, image, audio, and video synthesis; the rise of agentic autonomy transforming generative systems into goal-driven, autonomous entities; and the formalization of responsible AI governance through legally enforceable regulations. While this technological landscape has generated substantial economic impact, contemporary survey literature exhibits fundamental fragmentation, lacking theoretical integration, mathematical novelty, and quantitative complexity analysis. This article addresses these persistent gaps through a unified mathematical framework comprising 33 theorems. We establish that all generative models–Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), Transformers, and Diffusion models–solve equivalent optimization problems through different mathematical parameterizations: VAEs via Evidence Lower Bound (ELBO), GANs via Jensen-Shannon divergence minimization, Transformers via mutual information maximization, and Diffusion models via score matching. We derive convergence guarantees (<InlineEquation ID="IEq1"><EquationSource Format="TEX">\(O(\log T / T)\)</EquationSource></InlineEquation> for VAEs, exponential convergence for GANs), universal approximation bounds, and reconstruction error decompositions enabling systematic model diagnosis. Novel contributions include Adaptive Prior ELBO (1.8−10.5% improvement), Wasserstein Gradients (5–15% FID improvement), Optimal Attention Scaling (2–4 BLEU improvement), and Learned Noise Schedules (30–50% step reduction). We extend theoretical analysis to multimodal architectures (Product-of-Experts optimality, cross-modal information flow), agentic systems (reward-weighted generation), and application-specific constraints (medical imaging, clinical validity). Comprehensive complexity analysis provides formal <InlineEquation ID="IEq2"><EquationSource Format="TEX">\(O(\cdot )\)</EquationSource></InlineEquation> notation enabling principled model selection across computational budgets. A rigorous four-axis taxonomy classifies generative models by likelihood tractability, latent structure, training objective, and sampling strategy, facilitating systematic navigation of the design space and identification of unexplored architectural regions. Theory-to-practice translation is demonstrated through case studies in healthcare imaging, large language models, and enterprise AI systems, connecting rigorous theory with production-grade implementation. This work uniquely combines theoretical rigor with comprehensive scope and original algorithmic advances, establishing the unified mathematical and algorithmic foundation that contemporary generative AI requires.</p>

Exploring generative AI through core frameworks, emerging innovations, and applications

期刊论文
出版模式：

开放获取
发表日期：

2026年7月16日

查看全文

Sivarama Prasad Tera,
Ravikumar Chinthaginjala,
N. T. Priya,
Hyoungsuk Lee

摘要

Generative Artificial Intelligence has undergone rapid maturation between 2023 and 2025, driven by three converging paradigm shifts: the emergence of multimodal foundation models unifying text, image, audio, and video synthesis; the rise of agentic autonomy transforming generative systems into goal-driven, autonomous entities; and the formalization of responsible AI governance through legally enforceable regulations. While this technological landscape has generated substantial economic impact, contemporary survey literature exhibits fundamental fragmentation, lacking theoretical integration, mathematical novelty, and quantitative complexity analysis. This article addresses these persistent gaps through a unified mathematical framework comprising 33 theorems. We establish that all generative models–Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), Transformers, and Diffusion models–solve equivalent optimization problems through different mathematical parameterizations: VAEs via Evidence Lower Bound (ELBO), GANs via Jensen-Shannon divergence minimization, Transformers via mutual information maximization, and Diffusion models via score matching. We derive convergence guarantees (\(O(\log T / T)\) for VAEs, exponential convergence for GANs), universal approximation bounds, and reconstruction error decompositions enabling systematic model diagnosis. Novel contributions include Adaptive Prior ELBO (1.8−10.5% improvement), Wasserstein Gradients (5–15% FID improvement), Optimal Attention Scaling (2–4 BLEU improvement), and Learned Noise Schedules (30–50% step reduction). We extend theoretical analysis to multimodal architectures (Product-of-Experts optimality, cross-modal information flow), agentic systems (reward-weighted generation), and application-specific constraints (medical imaging, clinical validity). Comprehensive complexity analysis provides formal \(O(\cdot )\) notation enabling principled model selection across computational budgets. A rigorous four-axis taxonomy classifies generative models by likelihood tractability, latent structure, training objective, and sampling strategy, facilitating systematic navigation of the design space and identification of unexplored architectural regions. Theory-to-practice translation is demonstrated through case studies in healthcare imaging, large language models, and enterprise AI systems, connecting rigorous theory with production-grade implementation. This work uniquely combines theoretical rigor with comprehensive scope and original algorithmic advances, establishing the unified mathematical and algorithmic foundation that contemporary generative AI requires.