Large Language Models in Networked Prisoner’s Dilemma

Abstract

Large Language Models (LLMs) are increasingly used as autonomous agents capable of performing complex reasoning and decision-making tasks. This thesis investigates the strategic behavior of LLMs in repeated social dilemma environments, focusing on the Iterated Prisoner’s Dilemma (IPD). In particular, the study examines whether LLM-based agents can sustain cooperation, adapt to opponent behavior, and respond to contextual information when interacting in both pairwise and networked multi-agent settings. The experimental framework evaluates LLM agents against classical rule-based strategies as well as against other LLM agents. Beyond traditional two-player interactions, the work introduces networked environments where agents interact within structured graphs, including line, cycle, and star topologies. These settings enable the analysis of how network structure and agent position influence emergent strategic dynamics. To study the robustness of model behavior, the experiments incorporate controlled prompt variations and behavioral interventions. These include prompt reordering to measure prompt sensitivity through the Sensitive Index, as well as contextual interventions such as deception, soft deception, warning, and uncertainty prompts. These manipulations allow us to examine the extent to which LLM agents rely on textual cues versus observed interaction history when making decisions. The results reveal that LLM strategic behavior varies significantly across model families and scales. While cooperation can emerge under favorable conditions, it remains sensitive to prompt framing, opponent strategies, and network topology. These findings highlight both the potential and the limitations of current LLMs as decision-making agents in multi-agent environments, and contribute to the broader understanding of LLM behavior in dynamic strategic settings