What's this about?
RL post-training gets studied in isolation from the pretraining that precedes it; this paper connects them, asking how pretraining choices (model size, data composition) determine what RL can later elicit — when RL unlocks latent ability vs. teaches something new (HF 5↑).
What exactly is the pretraining/post-training split?
Pretraining is the massive read-everything phase where a model absorbs language and knowledge; post-training (RL, instruction tuning) is the shorter phase shaping it into a useful, reasoning assistant. A live scientific debate: does RL create reasoning ability or merely surface what pretraining already built? The answer decides where labs should spend compute.
Why does it matter?
If reasoning is mostly latent in pretraining, RL budgets are being overrated; if RL genuinely adds capability, scaling it (see LongStraw above) is the path. Evidence connecting the two phases directly is scarce and valuable.