Paper finds RL gains can be predicted from pretraining loss, with optimal compute share around 20%–28%

rohanpaul_ai · x · 2026-07-26

How much RL helps can be predicted from the checkpoint you start with

This paper recreates the full LLM training pipeline in chess: pretraining 5M–1B parameter models on Lichess games, SFT on synthetic reasoning traces, then GRPO on puzzles with verifiable rewards. Across 36 pretraining/RL combinations, the authors find that two pretraining properties explain most of the post-RL gain:

Their fitted frontier suggests RL should take only a minority share of compute: about 20% at 50M parameters and 28% at 680M parameters, rising with budget but still not becoming dominant.

The mechanism matters too. RL is not just sharpening the SFT policy uniformly: on easy puzzles it amplifies already-preferred correct moves, while on hard puzzles it surfaces correct moves that were almost absent under SFT. That helps pass@1, but not pass@16, leading the authors to argue the right direction is to suppress wrong-mode amplification rather than simply “make the policy sharper.” They report the same pattern in a 1B math model as well.

Original post →

More from Research

Research channel →