How Persona Prompts Shape Outcomes in LLM-Controlled Buyer–Seller Price Negotiations

Sammanfattning

This dissertation explores how personality conditioning in large language models (LLMs) affects the dynamics and outcomes of autonomous trade negotiations. Building on behavioral economics and computational psychology, the study investigates whether persona prompts can alter negotiation behavior among simulated buyer and seller agents. A controlled multi-agent simulation framework was developed, conducting 15,000 negotiation sessions in Python using LLM-based agents conditioned by personality-specific prompts. Each interaction produced quantitative data on three outcome variables: (1) agreement rate, (2) final negotiated price difference (ΔPrice), and (3) number of negotiation rounds. Statistical analyses, including two-proportion z-tests and Welch’s t-tests, were applied to assess three hypotheses regarding behavioral variation. Results showed that persona conditioning significantly influenced negotiation style and process variables. Agents with an Agreeable persona achieved higher agreement rates, supporting the trait’s cooperative nature. Machiavellian agents engaged in longer, more complex negotiations but did not secure higher economic advantage, indicating diminishing returns of manipulative strategy. Rational personas negotiated more efficiently, reaching agreements faster with consistent outcomes. These findings confirm that linguistic personality conditioning can reliably produce distinct, trait-aligned behaviors in LLM-controlled negotiation agents. The study contributes to the emerging field of AI behavioral modeling by demonstrating that LLMs can serve as psychologically interpretable agents capable of simulating human-like decision dynamics in economic contexts. Beyond theoretical insights, the work highlights practical implications for designing adaptive negotiation systems, personality-aware trading agents, and virtual markets governed by AI-driven social interaction principles.