New chess study fits a pretraining–RL scaling law to optimize compute allocation
micahgoldblum · x · 2026-07-21
New paper: a joint pretraining–RL scaling law in chess
Researchers fit a pretraining–RL scaling law in a controlled chess testbed to ask a practical question: as compute grows, should it go into stronger pretraining or more RL?
The paper combines three pieces:
- Pretraining on human chess games to model move sequences.
- SFT on synthetic reasoning data to turn trajectories into search-tree-style token sequences.
- RL in a verifiable puzzle environment where moves can be checked against ground truth.
Their scaling analysis traces the optimal compute allocation across pretraining and RL. The figure also argues that RL is not just “more data” or simple sharpening: it appears to improve performance differently on easy vs. hard puzzles, with stronger tail discovery on difficult cases.
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