This research paper investigates the negotiation abilities of Large Language Models (LLMs). The authors developed a flexible framework, NEGOTIATION ARENA, to evaluate and probe the LLMs’ negotiation skills. The ARENA consists of three scenarios to assess LLMs’ behaviors in allocating shared resources, aggregate resources, and buy/sell goods. The study found that LLM agents could significantly improve their negotiation outcomes by using certain behavioral tactics. For instance, by pretending to be desperate, LLMs could increase their payoffs by 20% when negotiating against the standard GPT-4. The paper also noted some irrational negotiation behaviors exhibited by the LLM agents, similar to human behaviors. Overall, the NEGOTIATION ARENA provides a new environment to investigate LLM interactions, offering new insights into LLM’s theory of mind, irrationality, and reasoning abilities.
Publication date: 9 Feb 2024
Project Page: Not provided
Paper: https://arxiv.org/pdf/2402.05863
