LLaDA-Instruct-JustGRPO-Math
This model is LLaDA-8B-Instruct fine-tuned with JustGRPO on MATH-500.
It was introduced in the paper The Flexibility Trap: Why Arbitrary Order Limits Reasoning Potential in Diffusion Language Models.
Method
JustGRPO is a minimalist RL approach for diffusion language models. Instead of complex diffusion-specific RL adaptations, we simply treat dLLMs as autoregressive models during training and apply standard GRPO. See our paper for details.
- Project Page: https://nzl-thu.github.io/the-flexibility-trap
- Paper: arXiv:2601.15165
- Code: https://github.com/LeapLabTHU/JustGRPO
Performance
MATH-500
| Sequence Length | 128 | 256 | 512 |
|---|---|---|---|
| Accuracy (%) | 39.0 | 45.1 | 45.2 |
Usage
For generation and evaluation, please refer to our GitHub repository.
Citation
@article{ni2026flexibility,
title={The Flexibility Trap: Why Arbitrary Order Limits Reasoning Potential in Diffusion Language Models},
author={Ni, Zanlin and Wang, Shenzhi and Yue, Yang and Yu, Tianyu and Zhao, Weilin and Hua, Yeguo and Chen, Tianyi and Song, Jun and Yu, Cheng and Zheng, Bo and Huang, Gao},
journal={arXiv preprint arXiv:2601.15165},
year={2026}
}