Paper fits scaling laws for agentic RL data across tasks, environments and trajectories

A paper on agentic reinforcement learning fits scaling laws along three data axes: tasks, environments and trajectories, according to Elvis Saravia, who describes it on X. The authors report that across model sizes, model families, domains, and both PPO and GRPO, the scaling exponents rank environments above tasks, and tasks above trajectories.

The authors also report that allocating a training budget with their recipe gives up to 1.7x compute efficiency compared with the common practice of generating more trajectories.