Paper

Understanding Reasoning from Pretraining to Post-Training

Jingyan Shen, Ang Li, Salman Rahman, Yifan Sun, Micah Goldblum, Matus Telgarsky, Pavel Izmailov

A controlled chess pipeline mirrors LLM pretraining, SFT, and RL to quantify how pretraining sets RL returns and how RL changes reasoning behavior.

arXiv ↗Jul 17, 2026Code ↗Dataset ↗

llm
nlp
reasoning
reinforcement learning
scaling laws

Multi-panel pipeline showing pretraining, SFT, RL method plus scaling law and mechanism analysis

What it is

Studies how pretraining quality and scale shape RL post-training for reasoning, using chess as a controlled LLM-style testbed and a math-domain transfer check. Instead of studying RL post-training alone, it models the full pretrain→SFT→RL pipeline. A controlled chess setup lets them link post-RL gains to pretraining loss/tokens and inspect whether RL sharpens existing policies or discovers rare correct moves.

Key results

36 pretraining-RL combinations across 5M-1B models were analyzed
20M frontier RL compute ratio rises from 5% to 32% as total compute grows
A 1B math LM pretrained on 10B-200B tokens shows the same RL scaling pattern

Better pretraining does not just raise the starting point for RL; it also increases how efficiently RL improves reasoning.

Worth your time, not all of your time.

We track new ML work across papers, code, and data, then keep only the pieces worth a closer look. Selected, not collected.

What it is

Key results

One useful ML email a week.

More from Jul 20 – Jul 26

Worth your time, not all of your time.