A learnable control block trained with GRPO to select token generation order in multimodal masked diffusion models improves text-to-image alignment and multimodal understanding over logit-based heuristics.
ArXiv , year=
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Reinforcing the Generation Order of Multimodal Masked Diffusion Models
A learnable control block trained with GRPO to select token generation order in multimodal masked diffusion models improves text-to-image alignment and multimodal understanding over logit-based heuristics.