xx18/Composition-RL-4B-Physics_Math

Composition-RL-8B

This repository contains the Composition-RL-8B model, developed as part of the research presented in the paper Composition-RL: Compose Your Verifiable Prompts for Reinforcement Learning of Large Language Models.

Model Description

Composition-RL is a data-efficient Reinforcement Learning with Verifiable Rewards (RLVR) approach designed to improve the reasoning capabilities of Large Language Models. It addresses the issue of "too-easy" prompts (pass-rate = 1) by automatically composing multiple verifiable problems into a single, harder verifiable prompt. This ensures the model continues to receive informative training signals throughout the RL process.

• Initial Model: Qwen3-8b-Base
• Training Dataset: MATH-Composition-199K
• Task: Mathematical Reasoning
• Paper: arXiv:2602.12036
• Code: GitHub - Composition-RL

Performance

As detailed in the paper, Composition-RL-8B consistently improves reasoning capability over RL trained on original, non-compositional datasets across various benchmarks.

Citation

If you find this work helpful, 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}
}