Active Generative Agents Lab

Intelligent biological agents act to understand the world, not merely react to it. In contrast, modern artificial systems largely remain passive learners that struggle with efficiency, robustness, and reliability in open-ended settings. Recent progress in large-scale generative models and reinforcement learning makes it possible to revisit this gap with new tools, yet a principled integration remains missing. The Active Generative Agents Lab at the UNIST Graduate School of AI develops a unified framework in which generative modeling and reinforcement learning jointly enable agents to act, explore, and learn in the real world. Generative modeling and reinforcement learning admit a common formulation under probabilistic learning over trajectories. This perspective enables new algorithmic designs that combine likelihood-based training with reward optimization, improving stability and credit assignment. We develop principled methods across energy-based models, diffusion models in continuous and discrete domains, and autoregressive models. Generative models can act as adaptive, agent-conditioned simulators rather than static data generators. We construct closed-loop simulation frameworks in which the data distribution evolves in response to the agent’s uncertainty, performance, and objectives. The agent is trained along an automatically generated curriculum that consists of targeted, information-rich experiences. Classical approaches to decision-making under uncertainty, such as Bayesian optimization and multi-armed bandits, provide theoretically optimal strategies but do not scale to the high-dimensional, structured settings encountered by modern foundation models. Our goal is to equip parametric policies, including pretrained foundation models, with the ability to perform