Pretrain-RL study finds the optimal RL share rises from 20% to 30% as models scale
Pavel_Izmailov · x · 2026-07-21
A pretrain-RL paper finds the optimal RL share rises with model size
A new paper on chess-based reasoning studies reports that the compute-optimal RL share grows from about 20% at 50M parameters to 30% at 700M.
- On easy puzzles, RL mainly reinforces already strong top-k actions from the SFT policy.
- On hard puzzles, RL can surface actions from the distribution tail — sometimes good, sometimes bad.
- The authors build a controlled chess testbed spanning pretraining, SFT, and RL, and find a scaling law linking pretraining and post-training performance.
- They also test a 1B math model and see the same pattern: longer pretraining checkpoints reach higher post-RL performance and improve faster under RL.
The paper’s broader point is that RL does not just sharpen the supervised policy; it can reveal useful moves that were nearly absent under SFT.
Related event: New Research Proposes Joint Scaling Law for Pretraining and RL(18 posts)→
More from Research
- Skyfall GS Uses Flux to Refine Gaussian Splatting, Accepted at ECCV 2026 — ducha_aiki · 2026-09-11
- Could 10k agents discover learning methods beyond backprop, or just tweak existing ones? — SeunghyunSEO7 · 2026-09-11
- Apodex Launches TRACES, First Benchmark for Evaluating 'Discoverative AI' on Real-World Problems — Faheem_uh · 2026-09-11
- TRACES grades the process, not the answer: six-dimension eval for open-ended AI science — Faheem_uh · 2026-09-11
- Apodex launches TRACES, a benchmark grading AI on open-ended discovery instead of known answers — Faheem_uh · 2026-09-11
- Cognition's SWE-2 uses a KKT duality argument in RL to shift the effort Pareto curve — YouJiacheng · 2026-09-11