LLMs Model Counterparties but Fail at Strategic Bargaining

A recent investigation published on arXiv (2605.16575) explores the ability of large language model (LLM) agents to engage in strategic negotiations within a structured multi-attribute bargaining framework. The findings indicate that while current LLMs can effectively understand a counterpart's preferences early in their reasoning processes, this insight does not consistently lead to better results for the informed party. Analysis at the turn level shows that agents often react based on perceived counterparty values but struggle to align these actions with their own valuable attributes. Typically, sellers exhibit greater flexibility, and under conditions of asymmetric information, the informed party frequently makes less advantageous concessions. This research underscores a disconnect between preference modeling and effective strategic negotiation in LLMs.