Webs of Power: Unveiling Autocratic Elite Networks and Their Influence on Leader Constraints Using Large Language Models (LLMs)

This paper is part of a larger project that demonstrates LLMs' utility in unveiling elite networks that shape autocracies’ behavior in foreign and security policy. The literature of international relations and comparative politics has demonstrated that networks of elites are essential in our understanding of countless processes and outcomes such as elite politics, leader survival, or foreign policy. The recent rise of autocrats’ prominence and aggression around the globe has only further highlighted elites’ importance in politics. Yet, data on such elite networks remain extremely scant. Due to the low accessibility of elites and limited information about their connections, text data such as news, books, or Wikipedia remain the largest sources of information about elites and their networks. However, the sheer amount of time, labor, and resources required to identify elite networks from vast text corpuses can overwhelm any team of researchers. This leads to a lack of systematic data on elite networks. Due to such data challenges, the study of autocratic factional dynamics has mostly been limited to case studies, which (despite their detailedness) can only offer country- or time-specific answers. Large-N comparative studies, which produce systematic answers to the above questions, remain difficult. To address this gap, this project adds to a growing literature that takes deeper dives into the elite groups. Specifically, I construct an original dataset that captures relationship networks of elites in autocratic government cabinets from 1990 to 2015. The data collection method is powered by recent advancements in large language models (LLMs). Using both manually collected data & teaching data from larger models like Llama 3.3-70B and Chat GPT-4, I fine-tune (train) Llama 3.1-8B to detect and extract relevant details from each elite’s biographical records (e.g., schools, job titles, organizations, job timeframe, and family members, and colleagues). The data collected by the finetuned model is then enhanced by human correction, before being used to build elite networks based on overlaps among elites’ working histories and family ties. In addition, this updated paper will also conduct comparisons between the fine-tuned Llama and other LLMs and various configurations of fine-tuning techniques, thereby investigating how various LLMs and computational procedures can aid social scientists in learning about elites. Contributions: The project offers a systemic analysis on network dynamics and power balances among autocratic elites (and their implications for policy behavior), which is ever more relevant given authoritarian regimes’ recent rise. Additionally, the dataset can be used to construct various network measurements for other elite-related research questions. This paper introduces a framework for fine-tuning LLMs, with human-in-the-loop supervision, to extract information about political entities with high accuracy and precision while minimizing labor and financial costs. More broadly, this framework can enable large-scale data collection of various entities and events besides elite politics, with relatively high accuracy and reduced resources.