Towards Pluralistic Alignment of LLMs: A Comprehensive Survey

Pluralistic alignment instils large language models (LLMs) with the capacity to reflect diverse human values and preferences. It offers safe deployment that avoids LLMs collapsing into monolithic perspectives while ensuring that these systems operate in accordance with ethical standards and safety protocols. In this survey, we provide a comprehensive analysis of pluralistic alignment for LLMs based on the different capabilities that models support. We review existing literature covering different methods, datasets, and evaluation metrics, highlighting their strengths, limitations, and open challenges. Finally, we identify and encourage future areas for research to increase the scope and impact of pluralistic modelling in contemporary AI systems.