Training Language Models for Bilateral Trade with Private Information

A team of researchers from Yale and other institutions has published a groundbreaking paper titled 'Training Language Models for Bilateral Trade with Private Information.' The study introduces a structured bargaining environment as a new benchmark for evaluating the strategic capabilities of large language models (LLMs). In this setup, AI agents negotiate via tool calls within an event-driven simulator, with binding offers separated from natural-language messages to enable automated, objective evaluation of performance.

In a major benchmark experiment, the researchers conducted a round-robin tournament involving five frontier models across 15,000 negotiations. The results revealed that the most effective strategies for maximizing surplus share and deal rate involved implementing price discrimination through sequential offers. Specifically, aggressive anchoring (starting with an extreme offer), calibrated concession, and temporal patience were the tactics that correlated most strongly with success. Conversely, accommodating strategies that conceded too quickly disabled price discrimination and yielded the worst outcomes.

The paper also details training experiments where the team fine-tuned open-weight Qwen3 models (8B and 14B parameters). They used a two-stage process: supervised fine-tuning (SFT) followed by reinforcement learning via Group Relative Policy Optimization (GRPO). SFT approximately doubled the model's surplus share but reduced deal rates, while the subsequent RL stage recovered deal rates at the cost of some surplus gains. A key finding was that SFT compressed surplus variation across different price tiers, and this behavior generalized to unseen opponents, suggesting the models learned proportional strategies rather than memorizing specific price points.