A Hybrid Generative Agent-Based Model of Nonviolent Resistance with Large Language Model-Enabled Activist
Microsoft Teams
Dr. Lance Fiondella
lfiondella@umassd.edu
https://teams.microsoft.com/meet/252984977868302?p=H2LcD3UFCOp6BL31w6
Thesis Advisor: Dr. Lance Fiondella - Electrical and Computer Engineering
Committee Members:
Dr. Gokhan Kul - Computer & Information Science
Dr. Long Jiao - Computer & Information Science
Abstract:
Past studies indicate that nonviolent resistance has achieved higher success rates in global political transformations than armed conflict. Agent-based modeling has been employed to simulate these movements, but prior models rely on rigid numerical thresholds and oversimplify agent interactions. This limits the capacity of simulations to analyze the strategic reasoning underlying human leadership. To address these limitations, this thesis presents a hybrid generative agent-based model that integrates large language models to guide the decision-making of activist agents while keeping other agent classes rule-based for computational feasibility.
The proposed model introduces three methodological enhancements. First, LLM-guided decision-making replaces rule-based activist movement. To prevent behavioral drift, a four-component structured system prompt anchors agent identity through role, persona, domain-specific world knowledge, and scenario context. Second, eight interdependent institutional support pillars replace a single abstract pillar type. Third, a linguistic transformation layer translates continuous numeric states of agents into semantic social descriptors, allowing activists to perform reasoning tasks, such as identifying and mobilizing highly aggrieved civilians. We drive activist agents with three open-weight Large Language Models (Llama-3.1-8B-Instruct, Ministral-8B-Instruct, Qwen3-8B) under Zero-Shot and Chain-of-Thought (CoT) prompting, yielding six experimental configurations.
Macro-level validation against the Nonviolent and Violent Campaigns and Outcomes (NAVCO) 1.2 dataset suggests all configurations approximate historical campaign success rates at low participation levels. Simulations indicate that high activist coordination and low fatality rates predict campaign success. However, micro-level analysis shows that CoT reasoning is required for behavioral realism. While Zero-Shot agents default to generic protests, CoT-enabled agents semantically evaluate institutional vulnerabilities and execute tactics aligned with pillar susceptibility: protest and persuasion, noncooperation, and intervention. These agents prioritize institutional and civilian outreach, actively avoiding confrontation with security forces. Micro-level validation against the Global Nonviolent Action Database (GNAD) suggests these CoT configurations approximate the historical prevalence of diverse tactical behaviors. While different foundational LLMs capture distinct aspects of real-world resistance, this study constitutes a simulated instance of Gandhian strategic principles. Ultimately, this thesis identifies CoT-enabled agents as a promising method for analyzing how specific leadership paradigms operate within diverse societies.
For further information, please contact Dr. Lance Fiondella at lfiondella@umassd.edu.