Masked diffusion language models generate text by gradually denoising a sequence, but they leave a key decision open: which token positions should be revealed at each step. Many current methods answer this with hand-designed rules, or they force visibility to increase in a fixed, monotonic way. That may be simpler, but it is not necessarily optimal.
The new paper, TIPS, proposes a different route: a tiny inference-time policy that learns to choose which positions to make visible. Because the policy is small and applied only at inference, it can be added to existing models without retraining them. The authors argue that this learned selection can better adapt to the content being generated than static rules.
The paper is a single source, so there is no conflicting view to compare. The abstract does not report specific results, so the practical gains remain to be seen, but the idea points toward more flexible decoding for diffusion language models.