A chess testbed studies how to split compute between pretraining, SFT, and RL
Pavel_Izmailov · x · 2026-07-21
The paper introduces a chess-based testbed to study how compute should be split across pretraining, SFT, and RL.
- It mimics a normal LLM pipeline: pretrain on human games, then SFT on synthetic reasoning traces, then RL on chess puzzles.
- The goal is to study pretraining-to-post-training scaling on a tractable compute budget.
- The setup lets the authors test how PT loss predicts later RL performance, and how compute allocation changes with scale.
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