UBC PhD Student Aaron Dharna to Present Groundbreaking "Foundation Model Self-Play" Research at RL Conference 2025

Vancouver, BC – Aaron Dharna, a Ph.D. student at the University of British Columbia, has announced his upcoming presentation on "Foundation Model Self-Play" (FMSPs) at the Reinforcement Learning Conference (RLC) later this year. Dharna expressed his enthusiasm via a tweet, stating, > "I'm really excited to be presenting FMSPs at @RL_Conference later this year!"

The research, titled "Foundation Model Self-Play: Open-Ended Strategy Innovation via Foundation Models," explores a novel approach to enhancing reinforcement learning agents. FMSPs leverage the extensive knowledge and generative capabilities of large pre-trained foundation models to guide and improve the self-play process, which is a fundamental method for AI agents to learn and refine strategies by playing against themselves. This method aims to overcome limitations in traditional self-play by introducing more diverse and innovative strategies.

Dharna's work focuses on the intersection of reinforcement learning, meta-learning, and game AI, with a goal of developing AI systems capable of continuous discovery and adaptation. His paper on FMSPs is co-authored with Cong Lu from Google DeepMind and Jeff Clune, Dharna's advisor at the University of British Columbia and a Senior Research Advisor at DeepMind. This collaboration highlights the significance and potential impact of the research within the artificial intelligence community.

The Reinforcement Learning Conference is a prominent annual event that convenes leading researchers and practitioners to discuss advancements in the field. While specific dates and the exact location for RLC 2025 have not been widely publicized, the conference is renowned for showcasing cutting-edge research in areas such as multi-agent systems, deep reinforcement learning, and real-world AI applications. Dharna's presentation is anticipated to be a key highlight, given the growing interest in integrating large language models and foundation models with traditional AI paradigms.

The implications of FMSPs are substantial for the future of AI, particularly in domains that demand strategic innovation and open-ended problem-solving. By incorporating the vast knowledge embedded within foundation models, reinforcement learning agents may achieve more robust and creative solutions, potentially accelerating AI development in complex fields like robotics, advanced game playing, and scientific discovery, where the generation of novel behaviors and solutions is crucial.