What changed
A preregistered preprint compared dense process rewards with ordinary GRPO across 1,668 runs on six models while matching generated-token and GPU-hour budgets. In the tested setup, Monte Carlo continuations consumed 91% of generated tokens and none of four dense variants beat standard GRPO at equal compute.
Why it matters
The result challenges training studies that compare reward methods per update while ignoring the cost of generating the reward. It gives post-training teams a concrete evaluation rule: equalize tokens and time before calling a method better.
What to do
Re-run one proposed reward change at equal generated tokens and GPU-hours, publish learning curves and seeds, and compare against a tuned GRPO baseline before scaling the intervention.
Caveat
The evidence is a preprint on six relatively small model settings and specific dense-reward constructions. It does not prove that every process reward fails or that the reported gains transfer to frontier-scale training.