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