A publication dated April 18, 2026 compares pretraining and reinforcement learning (RL) scaling, highlighting compute costs and evaluation challenges.
Published on April 18, 2026, the piece compares scaling laws for pretraining and reinforcement learning (RL). It notes that RL compute includes sampling and policy updates, and can be measured in FLOPs or GPU hours. It also distinguishes extrapolation within a run from extrapolation across runs.
The piece identifies compute and evaluation choices as obstacles to comparing RL experiments and extrapolating their results. To verify the content and context of its claims, consult the original publication and check how it defines the measures and types of extrapolation; the available summary does not specify those choices.