A joint scaling law links RL performance to model size, pretraining tokens, and compute
Pavel_Izmailov · x · 2026-07-21
The paper proposes a joint scaling law for RL performance in terms of model size N, PT tokens T, and RL compute C.
- At high compute, RL performance is well predicted by PT loss.
- The slope of RL improvement is close to log-linear in PT tokens.
- The figure shows large sweeps across 36 combinations of model size, PT tokens, and RL compute, with more PT improving both pass@1 slope and final performance.
Related event: New Research Proposes Joint Scaling Law for Pretraining and RL(18 posts)→
More from Research
- MIT's Ziming Liu: AutoResearch doesn't generalize — blind training tricks cause 'forking' disasters — ZimingLiu11 · 2026-10-26
- Stanford Data-Centric AI Workshop reopens site, extends poster deadline — jure · 2026-10-26
- 'Vibe verification': ~60% of AI-written formal specs prove the wrong property — basedjensen · 2026-10-26
- Ito: a browser that detects browsing intent and auto-organizes tabs into Flows — andrewkuznet · 2026-10-26
- Cornell study says brains process language as linear sequences, challenging Chomsky's hierarchy — anirbanbandyo · 2026-10-26
- Hi3D V1.0 open-sourced with 2M training assets and 90% slicing success rate — kimmonismus · 2026-10-26