Optimizing Decoding Paths in Masked Diffusion Models by Quantifying Uncertainty
Peng Sun, Tao Lin et al.
Paper
Details
#MaskedDiffusion #UncertaintyQuantification #DeepLearningResearch

Self-Correcting Masked Diffusion Models Enabled by PRISM Plug‑In

PRISM adds a single extra inference pass to pretrained masked diffusion models, letting them re‑generate low‑quality tokens and improving Sudoku solution rates. Read more: getnews.me/self-correcting-masked-d... #prism #maskeddiffusion

RL Boosts Masked Diffusion Language Models, Cutting Decoding Steps

EOSER and ASS decoding tricks plus the CJ‑GRPO RL algorithm let LLaDA‑8B‑Instruct keep accuracy while cutting inference steps. The study was submitted in September 2025. Read more: getnews.me/rl-boosts-masked-diffusi... #maskeddiffusion #llada8b

New Study Shows Near‑ε‑Free Complexity for Masked Discrete Diffusion

Euler samplers achieve ε‑accuracy with ~O(d² ε⁻³ᐟ²) score evaluations in masked diffusion; MATU reaches O(d ln d·(1‑ε²)) complexity, removing the ln(1/ε) term. Read more: getnews.me/new-study-shows-near-e-f... #eulersampler #maskeddiffusion #matue

Lavida-O: Elastic Masked Diffusion Model Advances Multimodal AI

Lavida-O combines a generation transformer with a larger understanding transformer, enabling object grounding, editing and 1024‑pixel text‑to‑image synthesis. Read more: getnews.me/lavida-o-elastic-masked-... #lavidao #maskeddiffusion

Masked Diffusion Models Unified as Energy Minimization Framework

A new study frames masked diffusion models as energy‑minimization problems and proposes a Beta‑distribution mask schedule that cuts diffusion steps while keeping fidelity. Read more: getnews.me/masked-diffusion-models-... #maskeddiffusion #betaschedule