Scaling Inference-Time Computation via Opponent Simulation: Enabling Online Strategic Adaptation in Repeated Negotiation

A team of researchers has introduced a novel approach that enables large language models (LLMs) to adapt strategically during repeated negotiations without requiring retraining. The method, detailed in arXiv paper 2602.19309, addresses a critical limitation where current LLMs excel in single-agent or stationary environments but struggle in dynamic, multi-agent settings where opponents evolve their strategies.

The technique embeds principles from game theory—specifically smooth Fictitious Play (sFP)—into LLM inference through two key components. First, an auxiliary opponent model learns to imitate the opponent's time-averaged behavior through in-context learning. Second, the system enhances best-of-N sampling by simulating multiple negotiation scenarios against this opponent model. Empirical evaluations across different repeated negotiation games show the method delivers approximately 40% performance improvement over static baselines.

This breakthrough matters because it moves beyond the traditional paradigm of offline training or fine-tuning. Instead of preparing for worst-case scenarios, LLMs can now adapt online based on interaction feedback. The approach scales inference-time computation rather than model parameters, offering a more flexible and efficient path to strategic adaptation. This has significant implications for applications ranging from automated business negotiations to complex multi-agent simulations where real-time strategic thinking is essential.