Mingyu Cao · Alvaro H.C. Correia · Christos Louizos · Shiwei Liu · Lu Yin

University of Surrey · Qualcomm AI Research · ELLIS Institute Tübingen · Max Planck Institute for Intelligent Systems · Tübingen AI Center

📧 Contact: The code can be contacted at [email protected]

SOAR is a confidence-switched position beam search decoding strategy for diffusion language models. The core idea is:

When there are high-confidence tokens in the sequence, SOAR selects parallel decoding for these tokens; otherwise, it employs position beam search to expand the search space.

SOAR achieves improved decoding quality without sacrificing decoding speed, averaging results across GSM8K, MBPP, and HumanEval.

git clone https://github.com/duterscmy/SOAR.git
cd SOAR
pip install transformers==4.46.2 torch==2.5.1 accelerate==1.12.0

We evaluate SOAR using the lm-evaluation-harness from EleutherAI.

cd eval_llada8b
bash eval_soar_llada.sh

cd eval_dream7b
bash eval_soar_dream.sh

If you use SOAR in your research, please cite:

@misc{cao2026searchaccelerateconfidenceswitchedposition,
    title={Search or Accelerate: Confidence-Switched Position Beam Search for Diffusion Language Models}, 
    author={Mingyu Cao and Alvaro Correia and Christos Louizos and Shiwei Liu and Lu Yin},
    year={2026},
    eprint={2602.10953},
    archivePrefix={arXiv},
    primaryClass={cs.CL},
    url={https://arxiv.org/abs/2602.10953}, 
}

This implementation is based on the LLaDA and Dream repositories. We thank the teams for open-sourcing their models and code.