New paper finds a joint scaling law from pretraining through post-training
Pavel_Izmailov · x · 2026-07-21
New paper links pretraining and post-training with a unified scaling law
The paper, “Understanding Reasoning from Pretraining to Post-Training,” studies the full LLM training pipeline rather than treating pretraining and RL as separate stages.
Key claims from the post:
- the authors find a joint scaling law across pretraining and post-training
- they analyze how compute should be allocated across the pipeline
- they study what RL is doing to the policy
The attached figure summarizes three views:
- pretraining loss scaling over parameter counts
- RL reward scaling over RL steps
- a joint pretraining-RL frontier over total compute
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