Nebius Cut Rollout Collection Time by 66.9%: Here's What Changed
The New Stack, Tuesday, October 27th, 2026
Nebius cut RL rollout collection time by 66.9% by treating inference as part of the optimizer and rethinking sandboxes and staleness.
In reinforcement learning for agents, rollout generation often dominates each training step, and a few slow trajectories can stall the batch.
Nebius VP of AI Training Di Jin recommends measuring time to a trainable batch alongside throughput and treating the inference engine as part of the RL optimizer.
Allowing slightly stale weights lets generation and training overlap, with policy divergence corrected by truncated importance sampling, and a Terminal-Bench-based experiment cut collection time for a 256-trajectory batch by 66.9 percent.
Other advice includes scaling CPU and sandbox capacity apart from GPUs and separating infrastructure failures from model failures.