A Survey of Social Intelligence Dimensions in Large Language Models through Social Deduction Games

Social deduction games are increasingly used as testbeds to evaluate the social capabilities of large language models (LLMs), yielding a growing body of work across diverse games, models, and experimental settings. However, existing studies often pursue heterogeneous objectives without a clear conceptual framework, resulting in a fragmented understanding of which aspects of social intelligence are being assessed. To date, the literature lacks a focused survey that consolidates prior research within the specific context of social deduction games and clarifies the evaluation dimensions addressed. This work proposes a conceptual framework and presents a systematic review of research on LLMs in this domain, organizing prior studies from a social intelligence perspective grounded in the Social AI literature. By explicitly characterizing the objectives underlying existing work, we enable a structured comparison across approaches and identify recurring limitations and open challenges. This synthesis provides a principled lens for analyzing prior work and supports a more interpretable and systematic evaluation of social intelligence in LLMs within social deduction game settings.