LLM Negotiations and Social Dilemmas

In this one-hour talk from the Simons Institute, Aaron Courville from IVADO, Université de Montréal, and Mila explores the game-theoretic perspective of LLM agents interacting with humans and other agents in mixed-incentive environments. Discover how artificially intelligent agents are being integrated into human decision-making and why evaluating them solely on rewards achieved in static environments is insufficient. Learn about Courville's research on reinforcement learning training of agent policies in general sum games that aren't purely cooperative. Explore the novel "Advantage Alignment" approach—a family of algorithms derived from first principles that efficiently guides policy learning toward more cooperative and effective policies. The presentation concludes with insights on applying these methods to LLM and agent interactions, addressing the challenges of safety-guaranteed LLMs in negotiation and social dilemma contexts.