Reinforcement learning has become the standard way to sharpen a language model's reasoning after the main training run. Yet the two stages are usually studied apart, which leaves two basic questions open: how do pretraining decisions like model size and data change what RL can add, and what is RL really doing to the model.

These are hard questions to answer with ordinary language models. Pretraining corpora are enormous and uncontrolled, so when a behaviour appears afterwards there is no clean way to tell whether it came from pretraining or from RL. Running systematic sweeps across both stages costs more than most researchers can spend.

So the authors use chess as a controlled setting instead. The appeal is that you get to know exactly what went into the model, which is precisely what a web-scale corpus denies you. It is the same instinct as studying genetics in fruit flies rather than in people: not because flies are the interesting case, but because control is what makes the question answerable at all.