Generative Simulators

As machine learning systems become increasingly agentic, static datasets, hand-written environments, and fixed benchmarks are proving insufficient for training, evaluation, and governance. Agents deployed in real-world settings must reason under uncertainty, interact over long horizons, and continuously remember and adapt as their understanding of tasks, tools, and constraints evolves. We introduce Generative Simulators, adaptive environments that jointly co-generate tasks, world dynamics, and reward functions. We believe that generative simulators constitute the foundational infrastructure for self-adaptive world modeling, extending beyond reinforcement learning algorithms and beyond human-curated datasets.