Evolutionary Decoding

Escaping Confidence Trap: Evolutionary Decoding for Mathematical Reasoning in Diffusion LLMs

---

News

[TBA] The code will be released after the paper review results are available.

Overview

Diffusion large language models generate text through block-wise progressive unmasking. In mathematical reasoning, however, local token confidence can become misaligned with global correctness, leading to the diffusion confidence trap.

Evolutionary Decoding is a training-free test-time decoding framework that improves mathematical reasoning in diffusion LLMs through:

• Step-wise selection, which preserves useful numerical-symbolic tokens during progressive unmasking.
• Block-wise mutation, which introduces structured numerical and symbolic alternatives when a decoding trajectory becomes trapped.

Code Release

This repository currently contains the project page and paper assets.

The core implementation of Evolutionary Decoding is temporarily not public. Since the method is lightweight and relatively easy to reproduce from the paper, we will release the code after the paper review results are available.

Repository Structure

code/
├── README.md
└── docs/
    ├── index.html
    ├── styles.css
    ├── script.js
    └── assets/

Contact

For questions about the paper or implementation release, please contact the authors listed on the project page.

BibTeX

@article{sun2026escaping,
  title={Escaping Confidence Trap: Evolutionary Decoding for Mathematical Reasoning in Diffusion LLMs},
  author={Sun, Zhenhong and Zhao, Hanqing and Bian, Yatao and Tu, Rongcheng and Xie, Liuyue and Zhang, Xu and Wang, Jue and Modolo, Davide and Dong, Daoyi and Tao, Dacheng},
  journal={arXiv preprint},
  year={2026}
}