Prompt Sensitivity to Context in LLM Multi-Agent Decision-Making: a Prisoner’s Dilemma Case Study

This thesis investigates prompt sensitivity in LLM-based multi-agent decision-making using the Iterated Prisoner's Dilemma as a controlled framework. Two agents built on GPT-4o-mini and Ministral-8B were tested across three narrative framings, multiple payoff structures, opponent strategies, and game lengths. Results show that contextual framing meaningfully shapes strategic behavior, with both models exceeding typical human cooperation baselines while following no single deterministic strategy. Question-based analysis revealed limitations in numerical reasoning and game-theoretic awareness, highlighting the critical role of prompt design in LLM-based agent systems.