Chess scaling testbed finds more pretraining improves RL pass@1 and final performance
Pavel_Izmailov · x · 2026-07-21
Chess testbed shows pretraining tokens strongly shape RL scaling
The authors run 36 combinations across model sizes from 20M to 1B parameters, varying pretraining tokens and RL compute to study pretraining-to-post-training scaling on a controlled chess benchmark.
Key findings:
- More pretraining consistently improves the pass@1 vs. compute slope and final RL performance.
- pass@16 also benefits from more pretraining, but is mostly flat during RL.
- They further verify the trend on natural-language math reasoning with OLMo models.
- PT loss remains highly predictive of RL performance, and the slope improves as pretraining tokens increase.
- RL mostly sharpens easy-task behavior, while on harder tasks it can occasionally discover a low-probability tail solution.
They frame the chess setup as an academic testbed for studying scaling laws, data, and training methods with reasonable 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