DataFlex-RL

A controlled study of the data policies that RLVR pipelines layer on top of GRPO: which rollouts to keep, how to weight them, and which domains to mix. Peking University and Zhongguancun Academy compare 13 configurations across 12 matched seeds on Qwen2.5-7B-Base over 12 math, logic, and science benchmarks. Uniform GRPO adds 7.76 points of domain-balanced accuracy over the base checkpoint, and none of the eight rollout-selection or reweighting methods, nor the three adaptive domain mixers, beats uniform sampling with a paired 95% confidence interval that excludes zero. A negative result that bears on how much of the RL-scaling literature survives seed variance.