Beyond Confidence: Adaptive and Coherent Decoding for Diffusion Language Models

A novel decoding method for DLMs from a new perspective.

Project Page

This is a preliminary PyTorch implementation of our paper:

Beyond Confidence: Adaptive and Coherent Decoding for Diffusion Language Models.

If you have any problems, please email me (ck.ee@my.cityu.edu.hk).

1. Run inference_optimization_release.py and choose different decoding types, i.e., "ours" (CCD) or "dream" （entropy）)

2. We provide the Trip benchmark for your quick validation.

3. We will continue to provide full codes as soon as possible.