DAPO: An Open-Source LLM Reinforcement Learning System at Scale

TL;DR

DAPO presents an open reinforcement-learning system and a collection of training interventions aimed at improving large-scale reasoning optimization.

Why it matters

It moves discussion from a single optimizer name to the system-level choices—sampling, clipping, length handling, and reward stability—that often determine whether RL runs work.

Key findings

1. 01

   Large-scale RL performance depends on a bundle of interacting system and optimization choices.

2. 02

   Open recipes make ablation and independent verification more practical.

Scaling dimensions

Models, methods & benchmarks

Algorithms

DAPO, Policy optimization

Topics