New Research Proposes Joint Scaling Law for Pretraining and RL

A new study titled "Understanding Reasoning from Pretraining to Post-Training" introduces a controlled chess testbed to integrate the pretraining, SFT, and reinforcement learning (RL) of large models into a unified computable framework. This work provides quantitative answers to long-standing questions about how compute should be allocated between pretraining and post-training, and how RL actually alters models. The paper, code, and models have been open-sourced.

Key Details and Core Scaling Laws

Through 36 combinatorial experiments covering model sizes from 20M to 1B, the research team found that the optimal compute allocation shifts with the budget. At lower compute levels, priority should be given to pretraining. However, as models scale from 50M to 700M and the total budget increases, the optimal strategy gradually shifts towards RL, with the optimal RL compute proportion rising to about 30%. Furthermore, the number of pretraining tokens generally follows the Chinchilla scaling law. The authors also propose a joint scaling law: in high-compute regions, RL performance can be well predicted by pretraining loss, and this predictive relationship strengthens as pretraining tokens increase.

RL Mechanisms and Generalization

Regarding how RL works, the authors note that for simple tasks, RL primarily sharpens already decent strategies (sharpening the distribution). For harder tasks, it can sometimes discover tail-end solutions with very low probability in the original strategy. This means RL is not merely fine-tuning but can unearth rare reasoning patterns. @burnytech previously raised a similar industry question about whether RLVR reinforces existing capabilities or elicits new reasoning, which this study partially answers. Additionally, the authors extended their chess testbed conclusions to natural language math reasoning tasks using OLMo, verifying that the strong predictive power of pretraining loss on RL gains still holds.

2026-07-20 ~ 2026-07-21 · 18 related posts

Primary sources