Reinforcement Learning RL post-training is now a major lever for reasoning-centric LLMs, but unlike pre-training, it hasn't had predictive scaling rules. Teams pour tens of thousands of GPU-hours into runs without a principled way to estimate whether a recipe will keep improving with more compute. A new research from Meta, UT Austin, UCL, Berkeley, Harvard, and Periodic Labs provides a compute-performance framework -- validated over >400,000 GPU-hours -- that models RL progress with a sigmoidal curve and supplies a tested recipe, ScaleRL, that follows those predicted curves up to 100,000 GPU-hours.

Pre-training often fits power laws (loss vs compute). RL fine-tuning targets bounded metrics (e.g., pass rate/mean reward). The research team show sigmoidal fits to pass rate vs training compute are empirically more robust and stable than power-law fits, especially when you want to extrapolate from smaller runs to larger budgets. They exclude the very early, noisy regime (~first 1.5k GPU-hours) and fit the predictable portion that follows. The sigmoidal parameters have intuitive roles: one sets the asymptotic performance (ceiling), another the efficiency/exponent, and another the midpoint where gains are fastest.

Why that matters: After ~1-2k GPU-hours, you can fit the curve and forecast whether pushing to 10k-100k GPU-hours is worth it -- before you burn the budget. The research also shows power-law fits can produce misleading ceilings unless you only fit at very high compute, which defeats the purpose of early forecasting.

ScaleRL is not just new algorithm; it's a composition of choices that produced stable, extrapolatable scaling in the study:

The research also compares fitted curves for prevalent recipes (e.g., DeepSeek (GRPO), Qwen-2.5 (DAPO), Magistral, MiniMax-M1) and reports higher asymptotic performance and better compute efficiency for ScaleRL in their setup.

This work turns RL post-training from trial-and-error into forecastable engineering. It fits sigmoidal compute-performance curves (pass-rate vs. log compute) to predict returns and decide when to stop or scale. It also provides a concrete recipe, ScaleRL, that uses PipelineRL-style asynchronous generation/training, the CISPO loss, and FP32 logits for stability. The study reports >400,000 GPU-hours of experiments and a single-run extension to 100,000 GPU-hours. Results support a clean split: some choices raise the asymptote; others mainly improve compute efficiency. That separation helps teams prioritize ceiling-moving changes before tuning throughput knobs.