Composition-RL-8B

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.

This specific checkpoint is the 8B version, initialized from Qwen3-8B-Base and trained on the MATH-Composition-199K dataset.

Model Description

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.

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

Usage

For evaluation and data generation instructions, please refer to the official GitHub repository.

Citation

If you find this work helpful for your research, please consider citing:

@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 Rural 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}
}