Dream 7B: Diffusion Large Language Models

Fields

id

sources/arch-dream-7b

URL

https://arxiv.org/abs/2508.15487

Publisher

Ye et al. (arXiv)

Kind

paper

Published

2025-08-21

Accessed

2026-10-03

notes

Iterative denoising of a whole sequence rather than one token per forward pass; 'tunable quality-speed trade-offs' via the number of inference steps; AR-based LLM initialisation and context-adaptive token-level noise rescheduling.

Notes

Iterative denoising of a whole sequence rather than one token per forward pass; 'tunable quality-speed trade-offs' via the number of inference steps; AR-based LLM initialisation and context-adaptive token-level noise rescheduling.

Cross-references

This record references no other records.