Citation Bureau
Vol. I
No. 358
XIII SEPTEMBER MMXXVI
Reference

What is GRPO?

GRPO

GRPO is a policy-gradient reinforcement learning algorithm introduced by DeepSeek in the DeepSeekMath paper that eliminates the value (critic) network by estimating advantages from group-relative comparisons of sampled outputs. The thread the material tracks is how and why it displaced earlier approaches.

How it developed

  • May 2026 - Kyle Corbitt said there was an almost 10-year gap between PPO, which used a value model to judge importance on a token-by-token basis, and GRPO, which threw that away because it felt wrong.
  • May 2026 - Corbitt said that with tens of thousands of tokens in a reasoning trace, there is no way to decide which rare token matters most, so all rare tokens get upvoted.
  • May 2026 - Corbitt said GRPO took off not because it was a quantum leap over what came before, but because DeepSeek did the engineering work to scale it and released an actual artifact model.
  • Sep 2026 - Edward Hughes said weights are used to adjust the advantages during GRPO, applying upweighting or downweighting of behaviors at the turn level rather than the rollout level.

In the evidence

Every line below is attributed to a named speaker.

Worth quoting

Kyle Corbitt on GRPO abandoning per-token credit assignment, and it working anyway.

“This is like I said a very unsatisfying answer. And I think that's one of the reasons why there was like an almost 10-year gap between PPO that had this value model that tried to you know, determine on a token by token basis and like GRPO where it's like hey, we're just going to throw that all away cuz it feels wrong. It feels like it shouldn't work. In practice it does though.”
Kyle Corbitt · 1 May 2026
Contrarian take

GRPO discards per-token credit assignment entirely, upvoting all rare tokens equally regardless of which mattered, and still works in practice, defying prior intuitions from PPO.

“Now, you could ask the question well, it's like there could be many rare tokens. If you got like thousands, you know, tens of thousands of tokens in a reasoning trace, how do you decide which rare token is most important? And you don't. You just throw up your hands and you say all the rare tokens get up voted the same way.”
Kyle Corbitt · 1 May 2026
Contrarian take

GRPO became dominant not due to algorithmic novelty but because DeepSeek did the scaling engineering and shipped a working artifact model that validated it.

“I think the reason GRPO specifically like that algorithm and that acronym like, you know, very concretely took off was not necessarily because it was like a big quantum leap on what came before. It was because DeepSeek did a lot of engineering work around actually scaling it and released an actual artifact model that worked really well with it.”
Kyle Corbitt · 1 May 2026
Best explained

Turn-level credit assignment via GRPO advantage weighting is crucial for stable training on long-horizon tasks, because upweighting or downweighting behaviors at the turn level rather than the rollout level avoids credit misassignment.

“We use these weights to adjust the advantages during GRPO. So what we're really saying is that when you're upweing or downweing the behaviors we want to do that on a turn level doing GRPO rather than on a rollout level.”
Edward Hughes · 11 Sep 2026
Citation Bureau · reference note, compiled from attributed expert discussion. Last updated 2026-09-13.