New paper reveals GRPO, Dr. GRPO, and DAPO are the same mechanism
Popular LLM reasoning training methods all control one number: answer disagreement.
A new paper from Yong Yi Bay and Kathleen A. Yearick (arXiv:2607.00152) reveals a surprising unity behind three of the most popular methods for training language models to reason: GRPO, Dr. GRPO, and DAPO. While each was previously presented as its own fix, the authors prove they are all three operations on a single number: the standard deviation of answer correctness across sampled responses. This metric measures how much the model's answers disagree for a given prompt. When the group is evenly split between correct and incorrect, the standard deviation is at its maximum, and the training update is largest. When all answers agree, the update is zero. GRPO divides by this number, Dr. GRPO drops the division, and DAPO discards groups where the standard deviation is zero. The group-standard-deviation identity shows these are not separate tricks but one adjustable dial that controls where and how strongly learning happens.
The paper validates this insight using the large real-world difficulty dataset Big-Math and in a controlled training run. The implications are practical: training efficiency can be improved by focusing compute on problems with high disagreement (split answers) rather than unanimous ones. This unified view also reveals which problems deserve the most weight and how many sample attempts each needs. For practitioners, this means simpler hyperparameter tuning and a clearer understanding of why certain training configurations work. The paper provides code and data, making it easy for teams to apply the identity to their own reasoning models and potentially reduce training costs by up to 50% on unanimous groups.
- GRPO, Dr. GRPO, and DAPO are all manipulating the group standard deviation of answer correctness—not separate tricks.
- Split answers (50/50 correct/incorrect) produce the largest training updates; unanimous answers contribute zero learning.
- Validated on the Big-Math difficulty dataset and controlled training runs; code and data are available open-source.
Why It Matters
Unifies reasoning training methods into one dial, enabling more efficient LLM optimization and reduced compute waste.