Yu et al. (2025) — ARXIV-2503.14476
Key takeaways
- The stated motivation is reproducibility rather than novelty: the technical details of state-of-the-art reasoning models are concealed — the paper names the o1 blog post and the R1 report — so the community cannot reproduce the RL results even when the models are open.
- DAPO — Decoupled Clip and Dynamic Sampling Policy Optimization — is presented with the four techniques the authors say make large-scale LLM RL work, rather than as a single trick.
- 50 points on AIME 2024 from a Qwen2.5-32B base.
- Training code (on verl) and a curated dataset released with it, which is the part that makes the four techniques checkable.
Standing in the anthology
One of the three GRPO variations the record now carries, each fixing a named defect: the clipping and sampling behaviour here, the length bias in LIT-167, and the token-level importance ratio in LIT-180.
Filed as a note rather than a practice. The record’s RL practice is SOTA-129’s three-stage recipe, sourced to Olmo 3; whether GRPO’s successors warrant a practice of their own depends on whether the anthology wants to carry RL algorithm choice as practice, and nothing does yet. What this supplies meanwhile is the observation that “we ran GRPO” is under-determined: three papers within a year each found a different defect in it worth naming.