We introduce LLaDA (Large Language Diffusion with mAsking), a diffusion model with an unprecedented 8B scale, trained entirely from scratch, rivaling LLaMA3 8B in performance.
The LLaDA-8B-Base and LLaDA-8B-Instruct are uploaded
in Huggingface. Please first install transformers==4.38.2 and employ the transformers to load.
from transformers import AutoModel, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained('GSAI-ML/LLaDA-8B-Base', trust_remote_code=True)
model = AutoModel.from_pretrained('GSAI-ML/LLaDA-8B-Base', trust_remote_code=True, torch_dtype=torch.bfloat16)
We provide get_log_likelihood() and generate() functions in get_log_likelihood.py
and generate.py respectively, for conditional likelihood evaluation and conditional generation.
You can directly run python chat.py to have multi-round conversations with LLaDA-8B-Instruct.
In addition, please refer to our paper and GUIDELINES.md for more details about the inference methods.
We will not provide the training framework and data as most open-source LLMs do.
However, the pre-training and Supervised Fine-Tuning of LLaDA are straightforward. If you have a codebase for training an autoregressive model, you can modify it to adapt to LLaDA with just a few lines of code.
We provide guidelines for the pre-training and SFT of LLaDA in GUIDELINES.md. You can also refer to SMDM, which has a similar training process to LLaDA and has open-sourced the training framework.
We use two evaluation methods: conditional likelihood estimation and conditional generation. For the base model, conditional likelihood estimation is applied to specific metrics and conditional generation to the rest. For the Instruct model, conditional generation is used for all metrics.
We implement conditional likelihood estimation using the lm-evaluation-harness library, while conditional generation is performed with an internal library, as lm-evaluation-harness lacks support for certain metrics (i..e, HumanEval-FIM).
Please refer to Appendix B.5. of our paper for all evaluation details.
We provide the code for evaluation using the open-source library lm-evaluation-harness. To begin, please install lm_eval==0.4.5 and refer to evaluation/eval.sh for the specific commands.
Here, we address some common questions about LLaDA.
Please refer to GUIDELINES.md for the guidelines. You can also refer to SMDM, which follows the same training process as LLaDA and has open-sourced its code.
Our motivation is not to improve BERT, nor to apply image generation methods like MaskGIT to text. Our goal is to explore a theoretically complete language modeling approach — masked diffusion models. During this process, we simplified the approach and discovered that the loss function of masked diffusion models is related to the loss functions of BERT and MaskGIT. You can find our theoretical research process in Question 7.
Specifically, LLaDA employs a masking ratio that varies randomly between 0 and 1, while BERT uses a fixed ratio. This subtle difference has significant implications. The training objective of LLaDA is an upper bound on the negative log-likelihood of the model distribution, making LLaDA a generative model. This enables LLaDA to naturally perform in-context learning, instruction-following, and ensures Fisher consistency for scalability with large datasets and models. You can also find a direct answer to this question in Section 2.1 of our paper.
Network structure and probabilistic modeling are two distinct approaches that collectively form the foundation of language models. LLaDA, like GPT, adopts the Transformer architecture. The key difference lies in the probabilistic modeling approach: GPT utilizes an autoregressive next-token prediction method, while LLaDA employs a diffusion model for probabilistic modeling.
Currently, LLaDA's sampling speed is slower than the autoregressive baseline for three reasons:
- LLaDA samples with a fixed context length;
- LLaDA cannot yet leverage techniques like KV-Cache;
- LLaDA achieves optimal performance when the number of sampling steps equals the response length. Reducing the number of sampling steps leads to a decrease in performance, as detailed in Appendix B.4 and Appendix B.6 of our paper.
In this work, we aim to explore the upper limits of LLaDA's capabilities, challenging the assumption that the key LLM abilities are inherently tied to autoregressive models. We will continue to optimize its efficiency in the future. We believe this research approach is reasonable, as verifying the upper limits of diffusion language models' capabilities will provide us with more resources and sufficient motivation to optimize efficiency.
Recall the development of diffusion models for images, from DDPM to the Consistency model, where sampling speed accelerated nearly 1000 times over the course of 4 years. We believe there is significant room for optimization in LLaDA's sampling efficiency as well. Current solutions, including semi-autoregressive sampling (as detailed in GUIDELINES.md), can mitigate the fixed context length issue, and consistency distillation can reduce the number of sampling steps.
For details on the pre-training process of LLaDA, please refer to Section 2.2 of our paper. During the total pre-training on 2.3T tokens, we encountered a training crash (loss becoming NaN) only once at 1.2T tokens. Our solution was to resume checkpoint and reduce the learning rate from 4e-4 to 1e-4.
5. Why is the final answer "72" generated earlier than the intermediate calculation step (e.g., 12 × 4 = 48) in Tab4?
The mask predictor has successfully predicted the reasoning process. However, during the remasking process, the reasoning steps are masked out again. As shown in the figure below, the non-white background represents the model's generation process, while the white-background boxes indicate the predictions made by the mask predictor at each step. We adopt a randomly remasking strategy.
This is because our pre-training and SFT data were designed for training an autoregressive model, whereas LLaDA directly utilizes data that contains identity markers.
LLaDA is built upon our two prior works, RADD and SMDM.
RADD demonstrated that the training objective of LLaDA serves as an upper bound on the negative log-likelihood of the model’s distribution, a conclusion also supported by MD4 and MDLM. Furthermore, RADD was the first to theoretically prove that masked diffusion models do not require time t as an input to Transformer. This insight provides the theoretical justification for LLaDA’s unmodified use of the Transformer architecture. Lastly, RADD showed that the training objective of masked diffusion models is equivalent to that of any-order autoregressive models, offering valuable insights into how masked diffusion models can overcome the reversal curse.
SMDM introduces the first scaling law for masked diffusion models and demonstrates that, with the same model size and training data, masked diffusion models can achieve downstream benchmark results on par with those of autoregressive models. Additionally, SMDM presents a simple, unsupervised classifier-free guidance method that greatly improves downstream benchmark performance, which has been adopted by LLaDA.
@article{nie2025large,
title={Large Language Diffusion Models},
author={Nie, Shen and Zhu, Fengqi and You, Zebin and Zhang, Xiaolu and Ou, Jingyang and Hu, Jun and Zhou, Jun and Lin, Yankai and Wen, Ji-Rong and Li, Chongxuan},
journal={arXiv preprint arXiv:2502.09992},
year={2025}
}