A Comprehensive Survey on Safety and Alignment Methods for Next-Generation Artificial Intelligence Models

The rapid emergence of next-generation artificial intelligence models, particularly large language models, multimodal foundation models, and autonomous agentic systems, has fundamentally transformed the capabilities and societal impact of AI. Alongside these advances, concerns related to safety, controllability, and alignment with human values have become central challenges for both researchers and practitioners. Traditional reliability and robustness techniques, originally designed for narrow AI systems, are increasingly inadequate for addressing the complex behaviors, emergent properties, and large-scale deployment scenarios of modern AI models. As a result, a diverse body of research has recently focused on developing systematic approaches for AI safety and alignment, spanning data curation, training-time constraints, post-training controls, architectural safeguards, and human-in-the-loop alignment strategies. This survey provides a comprehensive and critical review of state-of-the-art methods for ensuring the safety and alignment of next-generation AI models. We propose a structured taxonomy of safety threats and misalignment risks across data, model, interaction, and system levels, and analyze how existing techniques address these challenges. Particular attention is given to the interplay between safety and alignment, highlighting their mutual dependencies and the limitations of treating them as isolated objectives. Through comparative analysis, we identify key trade-offs among scalability, robustness, interpretability, and practical deployability. Finally, we outline open challenges and promising research directions toward achieving trustworthy, controllable, and value-aligned AI systems suitable for real-world, high-stakes applications.

کلیدواژه ها:

AI Safety ، AI Alignment ، Large Language Models ، Foundation Models ، Trustworthy Artificial Intelligence