Curriculum RL

Curriculum reinforcement learning studies how to arrange or adapt training experiences so an agent can master tasks that are difficult to learn all at once. Methods include staged task schedules, self-paced sampling based on success or learning progress, and adaptive curricula that generate challenges through teachers or adversarial self-play; some also scaffold exploration or change reward objectives over time. These approaches are used both in language-model post-training for reasoning, coding, and tool use, and in robotics and control, with evaluations ranging from math and agent benchmarks to manipulation, locomotion, and navigation tasks. A central challenge is choosing examples that are hard enough to teach something but still learnable, while ensuring progress transfers across tasks rather than overfitting to a narrow curriculum.