Composition-RL-30B-A3B

This repository contains the Composition-RL-30B-A3B model, a 30B parameter Mixture-of-Experts (MoE) model presented in the paper Composition-RL: Compose Your Verifiable Prompts for Reinforcement Learning of Large Language Models.

Overview

Composition-RL is a data-efficient Reinforcement Learning with Verifiable Rewards (RLVR) approach. It addresses the challenge of "too-easy" prompts (where the model already achieves a 100% pass rate) by automatically composing multiple verifiable problems into a single, more complex compositional prompt. This ensures the model continues to receive informative training signals throughout the RL process.

Model Details

• Architecture: Mixture-of-Experts (MoE)
• Base Model: Qwen3-30b-a3b-Base
• Training Method: Composition-RL
• Training Dataset: MATH-Composition-199K

Resources

• Paper: Composition-RL: Compose Your Verifiable Prompts for Reinforcement Learning of Large Language Models
• Code: Official GitHub Repository
• Collection: Composition-RL Collection

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}
}