DAPO: ByteDance and Tsinghua Open-Source a Reinforcement Learning Stack for LLM Training

A team from ByteDance Seed and Tsinghua AIR open-sourced DAPO on September 21, 2026 — a reinforcement learning system designed for training large language models at scale. The project includes the algorithm, code infrastructure, and curated datasets needed to reproduce its results, hosted on GitHub with model checkpoints on Hugging Face.

Results on AIME 2024

On AIME 2024, a competition-grade math benchmark, DAPO scored 50 points using the Qwen2.5-32B base model. That matches or exceeds the previous state-of-the-art result from DeepSeek-R1-Zero-Qwen-32B while using roughly half the training steps.

Efficiency gains in RL training are scientifically significant because compute — not algorithm novelty alone — often determines who can reproduce frontier results. Halving steps at matched benchmark scores lowers the barrier for university labs and startups running serious math-reasoning experiments.

What the release includes

The team released two model checkpoints: the full DAPO model at 50 points and an earlier version without token-level policy gradient loss and dynamic sampling at 44 points. Both are on Hugging Face.

Reproducibility artifacts include:

• DAPO-Math-17k — curated math dataset
• Training scripts verified against the latest verl version
• Evaluation pipelines for AIME 2024

Researchers interested in extending DAPO can fork the codebase and swap base models, reward functions, or sampling strategies without rebuilding infrastructure from scratch.

Why RL for math reasoning matters

Reinforcement learning on verifiable rewards — where answers can be checked automatically — has become the dominant paradigm for improving LLM reasoning since DeepSeek-R1 demonstrated public appetite for open reproduction. DAPO enters that landscape as another open stack with competitive scores and documented ablations.

For science and engineering audiences, the interesting question is not only whether 50% on AIME is state of the art, but which algorithmic components drive the efficiency gain. The ablation checkpoint at 44 points isolates the contribution of token-level policy gradient loss and dynamic sampling.

Broader implications

Open RL infrastructure accelerates peer review in AI research. When ByteDance Seed and Tsinghua AIR publish code alongside checkpoints, other groups can validate claims, find bugs, and extend methods — the normal scientific cycle applied to a field that too often relied on press releases alone.

For educators and students, DAPO is a teaching resource. Running RL on LLMs in a classroom setting was impractical two years ago. Today, a well-resourced lab can reproduce published scores with open data and scripts — a shift that belongs in any curriculum covering modern machine learning systems.