The building sector accounts for over 40% of global energy consumption, with Heating, Ventilation, and Air Conditioning (HVAC) systems responsible for nearly 60% of this share. Improving HVAC efficiency while maintaining occupant comfort has therefore become a critical challenge for smart building management. Conventional control strategies, such as rule-based methods and Model Predictive Control (MPC), often fall short when dealing with dynamic, multi-zone environments. In response, recent advances in Artificial Intelligence (AI) have introduced new directions for HVAC prediction and control. This review systematically analyzes 15 recent studies (2023-2025), classified into three main categories: (i) Graph-SpatioTemporal Prediction (C1), focusing on graph neural networks combined with temporal modules for predicting temperature, CO?, occupancy, and energy demand; (ii) Multi-Agent Reinforcement Learning (C2), enabling adaptive and decentralized HVAC control across multiple zones and subsystems; and (iii) Representation & Contrastive Learning (C3), which enhances time-series representation to improve data efficiency and generalization. The synthesis highlights key achievements: high prediction accuracy from graph-temporal models, up to 40% energy savings using MARL, and improved robustness through contrastive learning. However, gaps remain, including the limited adoption of multi-task prediction, insufficient exploration of curriculum learning and policy distillation in MARL, and minimal integration of contrastive learning into HVAC applications. Looking ahead, the review outlines a 5-10 year roadmap, emphasizing hybrid multi-task models, curriculum MARL, contrastive-RL integration, cross-building transferability, federated learning, and the vision of autonomous, self-evolving HVAC systems. By providing a comprehensive mapping of the state of the art and future opportunities, this review aims to guide researchers and practitioners toward developing AI-based HVAC solutions that are more efficient, adaptive, and occupant-centered.