Skip to content

HJ-Young/WavefrontDiffusion

Folders and files

NameName
Last commit message
Last commit date

Latest commit

ย 

History

6 Commits
ย 
ย 
ย 
ย 

Repository files navigation

WavefrontDiffusion: Dynamic Decoding Schedule for Improved Reasoning

License: MIT Paper

Official implementation of the paper "WavefrontDiffusion: Dynamic Decoding Schedule for Improved Reasoning" (ICLR 2026).


๐Ÿ’ก Overview

WavefrontDiffusion is a training-free dynamic scheduling strategy for Diffusion Language Models (DLMs).

Traditional schedules like Standard Diffusion suffer from premature end-of-sequence predictions, while BlockDiffusion breaks coherent semantic units due to rigid boundaries. WavefrontDiffusion addresses these by:

  • Adaptive Expansion: Dynamically expanding a "wavefront" of active tokens from finalized positions.
  • Context Completeness: Ensuring tokens are only denoised when surrounded by sufficient contextual information.
  • Compute Parity: Matching the computational cost (FLOPs) of block-based methods while delivering higher quality.

๐Ÿš€ Main Results

WavefrontDiffusion achieves State-of-the-Art performance across reasoning and code generation benchmarks:

Model GSM8K (Acc) MATH (Acc) HumanEval (Pass@1) BBH (Acc)
LLaDA-8B (Block) 80.74 40.62 45.73 43.23
LLaDA-8B (Wavefront) 82.03 41.04 47.56 44.30

All results obtained with a fixed budget of T=1024 steps.

๐Ÿ› ๏ธ Installation

# Clone the repository
git clone [https://github.com/HJ-Young/WavefrontDiffusion.git](https://github.com/HJ-Young/WavefrontDiffusion.git)
cd WavefrontDiffusion

# Environment setup (Python 3.9+, PyTorch 2.7.0+)
pip install -r requirements.txt

About

Official implementation of the paper "WavefrontDiffusion: Dynamic Decoding Schedule for Improved Reasoning" (ICLR 2026).

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors