LIDS Seminar: Lillian Ratliff 

"Strategic Uncertainty as Structure: Game-Theoretic Approaches to Scalable and Robust AI"

Modern AI systems learn through feedback from other agents, environments, and evaluators—not just static data. This raises fundamental challenges: scaling learning without expensive supervision, ensuring robustness under multi-agent interaction, and maintaining reliable behavior under endogenous and exogenous shifts. In this talk, I argue that game-theoretic abstractions offer a rich, comprehensive framework for addressing these challenges—both by exploiting natural interaction patterns (e.g., learner–judge, teacher–student) and by introducing new structure that leverages uncertainty in strategic interactions to shape dynamics, refine learning, and enable predictable and robust outcomes. During training, this perspective supports scalable auto-supervised learning and structured exploration of high-dimensional parameter spaces, reducing reliance on costly human annotation. At deployment, this perspective provides a principled approach to multi-agent interaction, enabling control over equilibrium selection, resilience to cross-play, and robustness to distribution shift and strategic uncertainty. I will illustrate these ideas through recent work on fine-tuning large language models via equilibrium-based formulations with tunable uncertainty tolerance, and on multi-agent reinforcement learning with structured objectives modulated by optimism and risk—revealing a shared foundation for cognitive and embodied AI systems. Along the way, I will highlight how interaction structure shapes learning outcomes, present algorithms with provable finite-sample guarantees, and time permitting, point to open problems in a more principled theory of learning in modern AI-enabled systems.