# How Well Can LLMs Negotiate? NegotiationArena Platform and Analysis

**Authors:** Federico Bianchi, Patrick John Chia, Mert Yuksekgonul, Jacopo Tagliabue, Dan Jurafsky, James Zou  
**arXiv:** [2402.05863](https://arxiv.org/abs/2402.05863)  
**Date:** 2024-02-08  
**Summary page:** https://www.summarizepaper.com/en/paper/2402.05863v1/

## Key points

- Negotiation plays a pivotal role in social interactions for navigating various aspects of life.
- Effective negotiation skills are crucial for haggling over prices and equitable resource distribution.
- Large language models (LLMs) need adept negotiation abilities as they are increasingly used as proxies for human agents.
- The study introduces NegotiationArena, a framework to assess and explore LLMs' negotiation prowess.
- Three distinct scenarios within NegotiationArena scrutinize LLM behaviors: ultimatum games, trading games, and price negotiations.
- LLM agents strategically leverage behavioral tactics to enhance negotiation outcomes significantly, such as adopting a facade of destitution and desperation to boost payoffs by 20%.
- The study reveals irrational negotiation behaviors in LLM agents similar to those observed in human negotiators.
- NegotiationArena provides a novel environment for exploring interactions among LLMs, offering insights into their theory of mind, irrational tendencies, and reasoning capabilities.

## Summary

In the realm of social interactions, negotiation plays a pivotal role in navigating various aspects of life. From haggling over prices to determining equitable resource distribution, effective negotiation skills are crucial. As the use of large language models (LLMs) increases as proxies for human agents, it becomes imperative for these LLM agents to possess adept negotiation abilities. This study delves into the efficacy of LLMs engaging in negotiations with one another and presents NegotiationArena - a versatile framework tailored to assess and explore their negotiation prowess. Within NegotiationArena, three distinct scenarios are crafted to scrutinize LLM behaviors across different negotiation contexts: allocating shared resources through ultimatum games, managing aggregate resources via trading games, and facilitating buy/sell transactions in price negotiations. The design allows for multifaceted dialogues between LLM agents over multiple turns, fostering complex negotiation dynamics. Noteworthy findings emerge as LLM agents strategically leverage behavioral tactics to significantly enhance their negotiation outcomes. For instance, adopting a facade of destitution and desperation enables LLMs to bolster their payoffs by 20% when pitted against the standard GPT-4 model. Additionally, this study sheds light on irrational negotiation behaviors exhibited by LLM agents - mirroring similar tendencies observed in human negotiators. By providing a novel environment for exploring interactions among LLMs, NegotiationArena serves as a conduit for unraveling insights into these models' theory of mind, irrational tendencies, and reasoning capabilities. Through meticulous analysis and experimentation within this platform, researchers gain deeper comprehension into the nuances of negotiating strategies employed by LLM agents - paving the way for enhanced understanding and optimization of their negotiation proficiencies.