He Kaiming's team proposes ELF: continuous embedding-space diffusion LMs beat discrete DLMs with fewer steps
alec_helbling · x · 2026-09-18
MIT researchers including Keya Hu, Yoon Kim, Jacob Andreas, and Kaiming He released ELF: Embedded Language Flows.
Core idea: While leading diffusion language models (DLMs) operate over discrete tokens, ELF runs continuous-time Flow Matching in the embedding space of existing models like T5, staying continuous until the final step where a shared-weight network maps to discrete tokens.
Key results:
- Staying in continuous space lets proven image-diffusion techniques like classifier-free guidance transfer directly
- ELF substantially outperforms both leading discrete and continuous DLMs with fewer sampling steps
- v2 adds distillation results: a flow-based LM distilled into a flow map model for one- and few-step generation
More from Research
- Carmack digs into RL Q-value overestimation, finds elegant fix in Relative Value Learning — ID_AA_Carmack · 2026-09-18
- Mantic AI beats all 676 humans in top forecasting tournament, raises $25M — tshevl · 2026-09-18
- AF3 paper: just increasing seeds lifted scores from 0.3 to 0.6, says sokrypton — sokrypton · 2026-09-18
- XGEN unveils world-sim prototype: JING interactive model plus DAO shared-world engine — anthara_ai · 2026-09-18
- Bespoke longevity LLMs match or beat 17 frontier AI models at assessing aging biology — EricTopol · 2026-09-18
- 53% of TMLR submissions desk-rejected: 3 of 7 authors couldn't answer basic questions about their papers — Dr_Atoosa · 2026-09-18