Understand this story
In workload experiments based on Terminal-Bench, Nebius reduced the fixed-weight collection time for 256-trajectory batches by 66.9% by adjusting routing, sandbox execution, and batch assembly. Di Jin, Vice President of AI Training at Nebius, pointed out that rollout generation often accounts for most of the time in a single training step, and teams should treat the inference engine as part of the post-training reinforcement learning optimizer, measuring efficiency by "trainable batch time" rather than tokens per second. He also emphasized that environment construction is the new data curation work, and that unmeasured policy staleness is the real problem. Currently, this result has been reported by The New Stack, with no further developments yet.
AI-generated from public reports
Developments
1 confirmed developments
Nebius experiment cuts 256-trajectory batch collection time by 66.9%
Source:The New Stack
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Oct 27
The New Stack
Nebius cuts rollout collection time by 66.9%; the bottleneck in agent reinforcement learning lies in inference
In Terminal-Bench-based workload experiments, Nebius reduced the fixed-weight collection time for a batch of 256 trajectories by 66.9% by adjusting routing, sandbox execution, and batch assembly. Di Jin, Vice President of AI Training at Nebius, noted that rollout generation often accounts for most of the time in a single training step, and that teams should treat the inference engine as part of the post-training RL optimizer and measure efficiency by "trainable batch time" rather than tokens per second. He also emphasized that environment construction is the new data curation work, and that unmeasured policy staleness is the real problem.