Composition-RL is a data-efficient Reinforcement Learning with Verifiable Rewards (RLVR) approach that addresses the challenge of "uninformative" easy prompts by automatically composing multiple verifiable problems into a single, harder verifiable prompt. This maintains informative training signals and improves reasoning performance across various model sizes.

Model Details

Developed by: Xin Xu, Clive Bai, Kai Yang, Tianhao Chen, Yangkun Chen, Weijie Liu, Hao Chen, Yang Wang, Saiyong Yang, Can Yang.
Initial Model: Qwen3-8B-Base.
Training Set: MATH-Composition-199K.
Paper: Composition-RL: Compose Your Verifiable Prompts for Reinforcement Learning of Large Language Models.
Code: GitHub Repository.

Citation

If you find this work helpful, please consider citing:

bibtex

@article{xu2026composition-rl,
  title={Composition-RL: Compose Your Verifiable Prompts for Reinforcement Learning of Large Language Models},
  author={Xu, Xin and Bai, Clive and Yang, Kai and Chen, Tianhao and Chen, Yangkun and Liu, Weijie and Chen, Hao and Wang, Yang and Yang, Saiyong and Yang, Can},
  journal={arXiv preprint arXiv:2602.12036},
  year={2026}
}

xx18/Composition-RL-14B

Composition-RL-8B

Model Details

Citation

xx18/
Composition-RL-14B