xx18/Composition-RL-4B

Composition-RL-8B

This repository contains the Composition-RL-8B model checkpoint, a version of Qwen3-8B-Base fine-tuned using the Composition-RL framework.

Composition-RL is a data-efficient Reinforcement Learning with Verifiable Rewards (RLVR) approach introduced in the paper: Composition-RL: Compose Your Verifiable Prompts for Reinforcement Learning of Large Language Models.

Overview

Composition-RL addresses the challenge of "too-easy" prompts (where the pass rate becomes 1 during training) by automatically composing multiple verifiable problems into a single, more complex, yet still verifiable prompt. This ensures that the model continues to receive informative training signals throughout the RL process, leading to improved reasoning capabilities across mathematical and scientific domains.

Model Details

• Base Model: Qwen3-8b-Base
• Training Set: MATH-Composition-199K
• Framework: Reinforcement Learning with Verifiable Rewards (RLVR)

Resources

• Paper: Composition-RL: Compose Your Verifiable Prompts for Reinforcement Learning of Large Language Models
• GitHub Repository: XinXU-USTC/Composition-RL
• Collection: Composition-RL Models and Datasets

Usage

For implementation details, including data generation and evaluation scripts, 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 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}
}