Large Language Diffusion with Ordered Unmasking (LLaDOU)

We introduce the Large Language Diffusion with Ordered Unmasking (LLaDOU), which is trained by reinforcing a new reasoning paradigm named the Diffusion Chain of Lateral Thought (DCoLT) for diffusion language models.

Compared to standard CoT, DCoLT is distinguished with several notable features:

• Bidirectional Reasoning: Allowing global refinement throughout generations with bidirectional self-attention masks.
• Format-Free Reasoning: No strict rule on grammatical correctness amid its intermediate steps of thought.
• Nonlinear Generation: Generating tokens at various positions in different steps.

News

• [Sep 2025] LLaDOU has been accepted by NeurIPS 2025. Congrats!
• [July 2025] Training code is provided!
• [May 2025] Released LLaDOU v0 Math and LLaDOU v0 Code models, their evaluation code and technique report.

Getting Started

Inference

import torch
from transformers import AutoTokenizer
from networks.lladou_v0 import LLaDOUModelLM, sample

tokenizer = AutoTokenizer.from_pretrained("models/LLaDOU-v0-Math")
model = LLaDOUModelLM.from_pretrained(
    pretrained_model_name_or_path="models/LLaDOU-v0-Math",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map="cuda",
)

problem = "What is the answer of 1+1?"
outputs = sample(
    model,
    problem,
    tokenizer,
    device="cuda",
)
response = outputs["responses"][0]
print(response)

Training

We provide an example to train LLaDOU on GSM8K dataset, feel free to change the configuration file!

accelerate launch --num_processes 8 --config_file configs/accelerate/fsdp.yaml train.py --config configs/gsm8k_64step_example.yaml

Evaluation

Prepare datasets as following:

├── datasets
│   ├── gsm8k
│   │   └── ...
│   ├── MATH
│   │   └── ...
│   ├── mbpp.jsonl
│   ├── mbpp_test.jsonl
│   └── HumanEval.jsonl.gz

• For GSM8K and MATH evaluation, please run scripts/eval_math.sh.
• For MBPP and HumanEval evaluation, please run scripts/eval_code.sh.

Evaluation Metrics

Citation

If this repository helps with your work, please consider giving a star and citation:

@inproceedings{huang2025reinforcing,
  title={Reinforcing the Diffusion Chain of Lateral Thought with Diffusion Language Models},
  author={Zemin Huang and Zhiyang Chen and Zijun Wang and Tiancheng Li and Guo-Jun Qi},
  journal={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
  year={2025}
}