Are LLMs Exploitable Negotiators ?

ICLR 2026 Conference Submission16284 Authors

Keywords: Large Language Models, Game Theory, Robustness, Multi-Agent Systems

TL;DR: We investigate the proficiency of LLMs as negotiators in multi-agent LLM seettings.

Abstract: Large Language Models (LLMs) have recently demonstrated surprising proficiency in strategic and social tasks, including bargaining and negotiation. Yet, their robustness in adversarial multi-agent interactions remains unclear. In this work, we study whether LLM-based agents are exploitable negotiators in the sense of game theory. To do so, we design a set of controlled game-theoretic environments — including auctions, markets, and public goods games — where Nash equilibria are analytically computable. These testbeds allow us to evaluate exploitability by comparing LLM outcomes against equilibrium predictions and against rational and adversarial opponents. Across multiple settings, we find that LLM negotiators systematically deviate from equilibrium: they tend to over-concede, are vulnerable to anchoring strategies, and often produce inefficient outcomes. Our results highlight that while LLMs can negotiate fluently, they remain strategically exploitable — raising concerns for their use in real-world interactions and opportunities for improving robustness through adversarial training and self-play.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 16284