Composition-RL is a data-efficient Reinforcement Learning with Verifiable Rewards (RLVR) approach that addresses the scarcity of informative training signals by automatically composing multiple verifiable problems into a single, harder compositional prompt.
As training progresses in RLVR, models often master "easy" prompts, resulting in a pass rate of 1 and reducing effective learning. Composition-RL mitigates this by creating new, complex, yet verifiable questions from existing data, maintaining a high level of difficulty and informative signals throughout training.
For evaluation and data generation instructions, please refer to the official
GitHub repository.
1@article{xu2026composition-rl,
2 title={Composition-RL: Compose Your Verifiable Prompts for Reinforcement Learning of Large Language Models},
3 author={Xu, Xin and Bai, Clive and Yang, Kai Rural and Chen, Tianhao and Chen, Yangkun and Liu, Weijie and Chen, Hao and Wang, Yang and Yang, Saiyong and Yang, Can},
4 journal={arXiv preprint arXiv:2602.12036},
5 year={2026}
6}
