New paper debunks RL batch scaling: 2.29x throughput doesn't mean faster learning

teortaxesTex · x · 2026-09-03

A new paper on batch scaling in LLM reinforcement learning targets the common habit of cranking up batch sizes just because GPUs look underutilized. The authors separate the systems side from the algorithmic side: autoregressive rollout generation is memory-bandwidth-bound at low concurrency, so a larger batch gives an immediate throughput jump (up to 2.29× on fixed hardware), but each step now burns far more samples, trading away learning efficiency. Their rule of thumb: a bigger batch only cuts real wall-clock time if the throughput gain outweighs the per-step sample cost.

Original post →

More from Infra

Infra channel →