This is a very interesting unified DLLM work! 🚀
I do have a question, though. In Section 5.3 (Synergy Across Various Tasks), you describe how unified generation and understanding can mutually benefit each other. I’m wondering whether there is any deeper or more quantitative evidence supporting this boosting effect.
For example, is there an analysis showing how reducing part of the generated-image samples would impact multimodal understanding performance? I feel such an ablation would help clarify the advantage of unified DLLMs compared with unified AR or AR + diffusion approaches.
Thanks again for the great work!
This is a very interesting unified DLLM work! 🚀
I do have a question, though. In
Section 5.3 (Synergy Across Various Tasks), you describe how unified generation and understanding can mutually benefit each other. I’m wondering whether there is any deeper or more quantitative evidence supporting this boosting effect.For example, is there an analysis showing how reducing part of the generated-image samples would impact multimodal understanding performance? I feel such an ablation would help clarify the advantage of unified DLLMs compared with unified AR or AR + diffusion approaches.
Thanks again for the great work!