DAPO

Decoupled Clip and Dynamic sAmpling Policy Optimization — an open-source LLM reinforcement-learning system at scale, from the SIA-Lab (the joint lab of ByteDance Seed and Tsinghua AIR, with HKU). DAPO hardens GRPO-style RLVR with four techniques: Clip-Higher (raising the upper importance-ratio clip to prevent entropy collapse), dynamic sampling (filtering prompt groups with all-correct or all-wrong rollouts, which carry no gradient signal), token-level policy gradient loss for long chain-of-thought, and overlong reward shaping. Result: 50 points on AIME 2024 with Qwen2.5-32B, beating DeepSeek-R1-Zero-Qwen-32B with 50% fewer training steps.

Fully open — algorithm, verl-based training code, and the curated DAPO-Math-17k dataset (17K competition problems with integer-normalized answers for unambiguous verification) — and one of the most-adopted post-training recipes of 2025–26: the DAPO loss and dataset are standard options across open RL stacks (e.g., the Marin project's RL framework and its Delphi reasoning-RL study).