The study investigates whether large language models can autonomously improve each other in a negotiation game through self-play, reflection, and criticism. Two LLMs (GPT and Claude) are used in a negotiating setting where they play the roles of a buyer and a seller, aiming to achieve the best possible deal. A third model acts as a critic to provide feedback and improve strategies. Here are the key details and findings of the study:

• Model Role Play: LLMs adopt various roles (buyer and seller) to negotiate deals, with feedback mechanisms to improve performance.
• Iterative Improvement: Models show differential ability to improve negotiation strategy based on feedback and the role they play.
• Evaluation of Outcomes: Deals get progressively better or worse based on the model’s ability to incorporate feedback over multiple rounds.
• Risk of Deal Breakdown: There is a higher chance of breaking the deal as negotiations progress through multiple rounds without reaching a satisfactory agreement.

This research sheds light on the potential of LLMs to not only understand but also improve upon complex interactive tasks autonomously. It opens up avenues for developing stronger AI agents that require minimal human intervention, which is crucial for applications requiring negotiation skills like business negotiations or diplomacy. Read more here.