UC Davis

Decentralized multi-agent decision-making systems have long underpinned the architecture of modern economies, real-world markets at all scales, and networked environments. These systems are characterized by multiple rational agents, each driven by distinct utility functions and local objectives, operating under information asymmetry and constrained, ego-centric perspectives. Within these shared environments, agents must navigate a balance between individual optimization and collective outcomes, often without any centralized signals. Such conditions give rise to distinctive challenges, including combinatorial computational complexity, non-stationarity induced by the concurrent learning and confounding adaptations of other agents, and the need for coordination under incomplete information and strategic uncertainty. These intricacies complicate the design of mechanisms and learning algorithms that reliably produce efficient, stable, and desirable collective outcomes. Recent advances in generative models have demonstrated promising capabilities of learning and capturing rich latent representations underlying complex target data distributions. In this thesis, we investigate the role of generative models as a foundational framework for decentralized multi-agent decision-making. First, we propose a representation learning framework that introduces auxiliary learning processes to embed information specific to multi-agent dynamics to supplement and enrich deep multi-agent reinforcement learning algorithms. We then study diffusion-based planning for maximizing key performance indicators within large-scale auto-bidding environments with varying underlying auction mechanisms and designs. Lastly, we experiment with the use of language models to represent reasoning capabilities within multi-agent systems, proposing a representation learning-based activation steering approach to align these models toward multi-agent objectives. Together, these contributions highlight the potential of generative modeling as a unifying paradigm for scalable, adaptive, and strategically grounded multi-agent decision-making.