Collaborative Research: SODA: Self-Organizing Decentralized Agents

Artificial intelligence systems increasingly operate as groups of interacting agents that divide work, share information, use tools, and make decisions over many steps. However, most current multi-agent artificial intelligence systems still rely on a central controller to assign tasks and coordinate behavior, creating scalability bottlenecks, single points of failure, and barriers to collaboration when agents have different capabilities or cannot fully share information. This project develops SODA, Self-Organizing Decentralized Agents, a framework for multi-agent artificial intelligence systems that coordinate through local peer-to-peer interactions rather than a central orchestrator. The project studies how groups of specialized agents assign roles, adapt to failures, preserve useful specialization, and avoid harmful collective behavior. The research has potential applications in healthcare, transportation, smart grids, smart cities, social simulation, and scientific discovery. It also supports education through new course materials, tutorials, workshops, and graduate student training in machine learning, reinforcement learning, optimization, information theory, and responsible artificial intelligence. The research is organized into three thrusts: Thrust 1 develops decentralized self-organization through peer-based task decomposition, role assignment, execution-time self-refinement, role enhancement, and information-theoretic tests of useful emergence; Thrust 2 builds a robust serverless training framework in which agents learn through peer-to-peer collaboration, recover from failures, reorganize dynamically, and share knowledge without harmful interference; and Thrust 3 introduces simulation-based training and inference, treating the role of the agents (e.g., the system prompts) as part of the latent dynamics of agent societies and using simulation-based inference, continual consensus, and multi-agent direct preference optimization to discover and fine-tune role configurations that yield desired collective outcomes. Together, the project establishes the theoretical and empirical foundations for resilient, specialized, and beneficial collective intelligence at scale. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.