Rollout
One sampled attempt at a task, start to finish, used to compute a reward during RL training.
A rollout is a single trajectory: the model is handed a task, acts until it finishes or gives up, and the whole attempt is scored by a verifier. RL algorithms like GRPO sample many rollouts per task and push the policy toward the ones that earned reward. Rollout throughput is often the real bottleneck in RL training.
Worked example: in reinforcement learning, one full run of the policy through an episode — take actions, observe rewards, record the trajectory — used to estimate how good the current policy is and to compute the gradient that improves it. Gotcha: rollouts are the expensive part of RL (each is a full generation/simulation), so sample-efficiency matters; and ‘on-policy’ methods must use rollouts from the CURRENT policy (old ones bias the update), which is why RL training alternates between generating fresh rollouts and updating on them.