Tag
This paper reinterprets masked diffusion language model decoding as continuous clean-state prediction, introducing a flow-based framework where tokens are updated continuously and asynchronously based on confidence, achieving 97% of LLaDA's performance with 25% of the decoding budget.