[Paper Review] How Well Can LLMs Negotiate? NegotiationArena Platform and Analysis

Federico Bianchi, Patrick John Chia+4 more|arXiv (Cornell University)|Feb 8, 2024

Corporate Governance and Law4 citations

TL;DR

This paper introduces NegotiationArena, an open-source platform for evaluating and probing LLM agents in multi-turn negotiation scenarios. It demonstrates that LLMs like GPT-4 can significantly improve payoffs through strategic behaviors—such as feigning desperation—while also exhibiting irrational biases like anchoring and numerosity effects.

ABSTRACT

Negotiation is the basis of social interactions; humans negotiate everything from the price of cars to how to share common resources. With rapidly growing interest in using large language models (LLMs) to act as agents on behalf of human users, such LLM agents would also need to be able to negotiate. In this paper, we study how well LLMs can negotiate with each other. We develop NegotiationArena: a flexible framework for evaluating and probing the negotiation abilities of LLM agents. We implemented three types of scenarios in NegotiationArena to assess LLM's behaviors in allocating shared resources (ultimatum games), aggregate resources (trading games) and buy/sell goods (price negotiations). Each scenario allows for multiple turns of flexible dialogues between LLM agents to allow for more complex negotiations. Interestingly, LLM agents can significantly boost their negotiation outcomes by employing certain behavioral tactics. For example, by pretending to be desolate and desperate, LLMs can improve their payoffs by 20\% when negotiating against the standard GPT-4. We also quantify irrational negotiation behaviors exhibited by the LLM agents, many of which also appear in humans. Together, \NegotiationArena offers a new environment to investigate LLM interactions, enabling new insights into LLM's theory of mind, irrationality, and reasoning abilities.

Motivation & Objective

• To develop a flexible, open-source framework for evaluating LLM agents in dynamic, multi-turn negotiation settings.
• To investigate how LLMs behave in real-world negotiation scenarios involving resource allocation, trading, and price bargaining.
• To identify strategic, social, and irrational behaviors that affect negotiation outcomes among LLM agents.
• To quantify the impact of behavioral tactics—such as pretending to be desperate—on negotiation success.
• To assess the rationality and theory of mind capabilities of state-of-the-art LLMs in competitive, interactive environments.

Proposed method

• Design and implement a modular platform, NegotiationArena, supporting three core negotiation scenarios: resource exchange, multi-turn ultimatum games, and price negotiations.
• Structure each scenario with defined initial resources, goals, turn limits, and acceptance conditions to enable controlled, repeatable experiments.
• Integrate LLM agents (GPT-4, GPT-3.5, Claude 2.1, Claude 2) into the negotiation flow using flexible, turn-based dialogue interactions.
• Collect and analyze negotiation trajectories, including proposal sequences, acceptance/rejection patterns, and final payoffs.
• Measure performance using quantitative metrics such as final payoff share and win rate across different strategies and resource levels.
• Conduct ablation studies to evaluate the impact of specific behaviors—like emotional appeals or insults—on negotiation outcomes.

Experimental results

Research questions

• RQ1 How do LLM agents perform in multi-turn, dynamic negotiation scenarios compared to static benchmark tasks?
• RQ2 What strategic behaviors—such as feigning desperation or aggression—improve an LLM agent’s payoff in negotiations?
• RQ3 To what extent do LLM agents exhibit irrational behaviors like anchoring bias or numerosity bias in negotiation settings?
• RQ4 How does the absolute amount of resources to be split affect the fairness and acceptance of offers in ultimatum-style negotiations?
• RQ5 How do different LLMs (e.g., GPT-4 vs. Claude 2.1) compare in negotiation performance and behavioral patterns?

Key findings

• GPT-4 consistently outperforms other LLMs in negotiation outcomes across all scenarios, demonstrating superior strategic reasoning and communication.
• Pretending to be desperate or desolate increased an LLM’s payoff by up to 20% when negotiating against standard GPT-4.
• LLM agents exhibit anchoring bias, where initial offers disproportionately influence final outcomes, even when irrational.
• As the total amount to split increases (e.g., from $100 to $10B), the proposer’s share rises significantly—reaching nearly 79% at $10B—indicating a numerosity bias.
• Agents frequently accept unfair offers when the absolute value is large, mirroring human irrational behavior observed in classical ultimatum games.
• All tested LLMs show vulnerabilities to social manipulation tactics, such as insults or emotional appeals, which can be exploited to improve negotiation outcomes.