The Illusion of Rationality: Tacit Bias and Strategic Dominance in Frontier LLM Negotiation Games

Large language models (LLMs) are increasingly being deployed as autonomous agents on behalf of institutions and individuals in economic, political, and social settings that involve negotiation. Yet this trend carries significant risks if their strategic behavior is not well understood. In this work, we revisit the NegotiationArena framework and run controlled simulation experiments on a diverse set of frontier LLMs across three multi turn bargaining games: Buyer Seller, Multi turn Ultimatum, and Resource Exchange. We ask whether improved general reasoning capabilities lead to rational, unbiased, and convergent negotiation strategies. Our results challenge this assumption. We find that models diverge into distinct, model specific strategic equilibria rather than converging to a unified optimal behavior. Moreover, strong numerical and semantic anchoring effects persist: initial offers are highly predictive of final agreements, and models consistently generate biased proposals by collapsing diverse internal valuations into rigid, generic price points. More concerningly, we observe dominance patterns in which some models systematically achieve higher payoffs than their counterparts. These findings underscore an urgent need to develop mechanisms to mitigate these issues before deploying such systems in real-world scenarios.

💡 Research Summary

This paper revisits the NegotiationArena framework to evaluate whether the latest frontier large language models (LLMs) exhibit more rational, unbiased, and convergent behavior in multi‑turn bargaining scenarios. The authors test six state‑of‑the‑art models—Google DeepMind’s Gemini 2.5 Pro and Flash, OpenAI’s GPT‑4.1, GPT‑4o, GPT‑4.1 mini, and Anthropic’s Claude 4.5 Sonnet—across three classic negotiation games: a bilateral Buyer‑Seller market, a multi‑turn Ultimatum game, and a Resource Exchange setting. All models are accessed via official APIs, prompted identically to the original NegotiationArena implementation, and run with a temperature of 0.7 to control stochastic variance.

Key findings overturn the naive expectation that higher general reasoning ability automatically yields game‑theoretic rationality. First, strong numerical and semantic anchoring effects persist: the initial offer heavily predicts the final agreement price for every model, indicating that early “anchors” dominate the internal valuation process. Second, instead of converging to a single Nash equilibrium, models diverge into distinct, model‑specific strategic equilibria—each model develops its own “strategic signature.” For example, Gemini 2.5 Pro consistently secures high payoffs both as buyer and seller, while GPT‑4.1 mini struggles to capture surplus in either role. Third, pairwise interactions reveal systematic dominance patterns: Gemini 2.5 Pro and GPT‑4o routinely achieve higher utilities against weaker counterparts, suggesting a predatory dynamic where larger, more capable models can exploit less capable agents. Fourth, classic cognitive biases documented in prior work (anchoring, over‑confidence, omission bias) remain evident, demonstrating that scaling does not erase these human‑like flaws.

The three games expose different failure modes. In the multi‑turn Ultimatum, overly aggressive proposers trigger high rejection rates, leading to zero‑payoff deadlocks. In the Resource Exchange game, models split between diversification strategies and single‑resource hoarding, further evidencing divergent strategic philosophies. Sensitivity analyses show that the crossover point where buyer and seller payoffs equal shifts with the width of the Zone of Possible Agreement (ZOPA), and this shift varies markedly across models.

Overall, the study argues that improving raw model size or benchmark scores is insufficient for safe deployment of LLM‑based negotiators. Instead, targeted mechanisms to mitigate anchoring, ensure fairness, and monitor dominance dynamics are required before these agents are entrusted with real‑world economic, political, or social negotiations. The authors call for systematic bias‑mitigation techniques, standardized fairness evaluation frameworks, and regulatory oversight to prevent emergent predatory behavior in multi‑agent AI ecosystems.