Dream-7B

Researchers have developed a new algorithm for constrained decoding in diffusion language models, enabling efficient sampling under constraints defined by finite automata. Tests on models like Dream-7…

Researchers at arXiv have introduced "infilling extraction," a new method for extracting training data from diffusion language models (DLMs) that uses arbitrary binary masks instead of relying solely …