ELF: Embedded Language Flows for Continuous Diffusion Language Modeling
Researchers have introduced Embedded Language Flows (ELF), a novel class of diffusion models designed to enhance language generation by operating primarily within continuous embedding space. Unlike current leading diffusion language models that handle discrete tokens, ELF utilizes continuous-time Flow Matching and only maps to discrete tokens at the final step via a shared-weight network. This approach allows for the seamless adaptation of established techniques from image-domain diffusion models, such as classifier-free guidance. Experimental results demonstrate that ELF significantly outperforms existing discrete and continuous diffusion language models, achieving superior generation quality with fewer sampling steps. The study suggests that ELF provides a promising pathway for developing effective continuous diffusion language models, bridging the gap between successful image generation techniques and natural language processing. The paper was submitted to arXiv in May 2026 by a team including Keya Hu, Linlu Qiu, and Kaiming He, highlighting advancements in computational linguistics and artificial intelligence.
Wire timeline
ELF: Embedded Language Flows for Continuous Diffusion Language Modeling
Researchers have introduced Embedded Language Flows (ELF), a novel class of diffusion models designed to enhance language generation by operating primarily within continuous embedding space. Unlike current leading diffusion language models that handle discrete tokens, ELF utilizes continuous-time Flow Matching and only maps to discrete tokens at the final step via a shared-weight network. This approach allows for the seamless adaptation of established techniques from image-domain diffusion models, such as classifier-free guidance. Experimental results demonstrate that ELF significantly outperforms existing discrete and continuous diffusion language models, achieving superior generation quality with fewer sampling steps. The study suggests that ELF provides a promising pathway for developing effective continuous diffusion language models, bridging the gap between successful image generation techniques and natural language processing. The paper was submitted to arXiv in May 2026 by a team including Keya Hu, Linlu Qiu, and Kaiming He, highlighting advancements in computational linguistics and artificial intelligence.
cs.AI updates on arXiv.org