Chess-based study ties pretraining loss to post-RL reasoning

A quiet result from a new arXiv paper is worth sitting with if you have any exposure to how frontier labs are choosing to split compute between pretraining and RL post-training. A team with authors from NYU, Modal Labs, UCLA, UIUC and Columbia set out to answer a question that is oddly under-studied given how much money is going into RL for reasoning: how do pretraining choices actually shape the returns to RL compute, and what does RL really do to the model.

Their move was to strip the question down. Instead of arguing over open-domain reasoning benchmarks where you cannot control the training data, they used chess as a testbed, pretraining language models from 5 million to 1 billion parameters on human chess games, running supervised fine-tuning on synthetic reasoning traces, then RL on chess puzzles with verifiable rewards. The headline result is that the post-RL performance at a given RL compute level is well-predicted from the pretraining loss, with the slope of the RL reward curves improving approximately linearly with the pretraining tokens.

That is a stronger claim than it sounds. It says the base model does not just help RL, it forecasts where RL will end up. The authors also report that RL does not simply sharpen the SFT policy: on easy puzzles it amplifies correct moves the SFT policy already preferred, while on hard puzzles it surfaces correct moves that were nearly absent under SFT. To check this was not a chess artifact, they trained a 1B model on math-domain text and saw the same pattern, with longer-pretrained checkpoints reaching higher post-RL performance and improving faster under RL.

The honest caveat is that this is a controlled study on chess plus one math-text transfer run at 1B parameters, orders of magnitude below the frontier. What the paper does not directly settle is whether the same tight relationship holds on open-ended reasoning, coding, or agentic tasks, or whether some amount of RL compute can eventually close a gap left by weak pretraining. Take the specifics as reported by the authors, not as a settled scaling law.

If it does generalize, the practical implication is clean: pretraining loss becomes a cheap early signal for whether an expensive RL campaign is worth booking, and the budget conversation swings back toward the base model before another post-training run gets funded.