1 paper · 1 filter
Xinyun Wang, Min Zhang, Sen Cui +4
Diffusion language models enable parallel token generation through block-wise decoding, but their irreversible commitments can lead to stagnation, where the reverse diffusion proce…