“RLVR offers a high ratio of capability improvement per dollar spent on compute, leading labs to reallocate compute resources from pretraining to RLVR runs.”