RLVR Trains LLMs for Strategic Multi-Buyer Negotiations

Negotiation is a complex strategic interaction, especially for a single seller dealing with multiple buyers, each having private budgets and valuations. While Large Language Models (LLMs) are proficient in language, they often fall short as economic decision-makers in such scenarios, tending to fixate on the highest current bid rather than strategically exploring the market for better opportunities. This research addresses this gap by proposing a specialized training approach called Reinforcement Learning from Verifiable Rewards (RLVR). RLVR anchors the reward function to objective economic outcomes, allowing the LLM-based seller to natively learn the strategic balance between market discovery and surplus extraction. The results demonstrate that the trained seller undergoes a multi-stage strategic evolution, mastering techniques like price anchoring and strategic probing to identify and close deals with high-value buyers. This leads to substantially higher surplus extraction compared to frontier models. Furthermore, the learned seller strategies prove robust and generalize effectively to unseen buyer negotiation styles and budget distributions, showcasing a significant advancement in applying AI to complex economic bargaining.