Back to feed
arXiv cs.LG·

Adaptive Order Policies for Masked Diffusion

Signal
72
Hype
15
In three linesMasked diffusion models: a lightweight policy network learns optimal token unmasking order. Loss reweighting by denoiser probabilities. Outperforms heuristics on order-sensitive tasks like combinatorics and proteins.
Read source
Your take?
PapersReasoningBenchmarks

Summary generated by Claude — human-verified