Hi,
Thank you for your effort on the survey paper, it's really helpful.
Our ICML2025 paper GraphGPT: Generative Pre-trained Graph Eulerian Transformer employ the Scheduled Masked Token Prediction (SMTP) pre-training on lossless serialized graph data, and show that it's better than NTP pre-training.
The SMTP is adopted for MaskGIT directly, and it is closely related discrete diffusion in language.
This might implies that discrete diffusion could be a hopeful candidate for unifying various modalities, including graph.
It would be great if you can have a look at our paper. Any feedback and questions are welcome.
Thank you.
Hi,
Thank you for your effort on the survey paper, it's really helpful.
Our ICML2025 paper GraphGPT: Generative Pre-trained Graph Eulerian Transformer employ the Scheduled Masked Token Prediction (SMTP) pre-training on lossless serialized graph data, and show that it's better than NTP pre-training.
The SMTP is adopted for MaskGIT directly, and it is closely related discrete diffusion in language.
This might implies that discrete diffusion could be a hopeful candidate for unifying various modalities, including graph.
It would be great if you can have a look at our paper. Any feedback and questions are welcome.
Thank you.