scaling laws for RL

In language modeling, scaling laws turned progress into a recipe: loss falls in a smooth, predictable power law as you grow model size, data, and compute, so you can plan capability in advance. Scaling laws for RL ask whether decision-making obeys the same kind of clean curve — does an agent's performance improve predictably as you pour in more parameters, more environment interaction, and more compute?

The picture is messier than in supervised learning. RL adds knobs language modeling does not have — exploration, the amount and diversity of environment interaction, the number of parallel actors, and the choice of algorithm — and the signal is a noisy, non-stationary reward rather than a clean loss. Still, researchers have found reasonably regular trends in places: model and data scaling in offline RL and sequence-model agents, and compute-versus-performance curves in large self-play and game-playing systems.

Getting RL scaling laws right would be transformative, letting labs forecast how much compute and interaction buys a target skill instead of guessing. But the open questions are serious: which axis (data, model, exploration) binds first, whether curves break or plateau, and how to even define the data axis when an agent generates its own experience. It connects directly to the broader study of neural scaling laws and to foundation models for RL.