The paper defines Informed Masking (IM) by scoring each valid completion token with the summed absolute change in neighboring still-masked positions’ top-1 confidence immediately after that token is revealed, then sampling a fixed-size mask without replacement with a softmax bias toward low-score downstream tokens. IM changes only mask selection; the host reward, mask count, diffusion-time schedule, and policy-gradient estimator remain unchanged. · CiteArk
The paper defines Informed Masking (IM) by scoring each valid completion token with the summed absolute change in neighboring still-masked positions’ top-1 confidence immediately after that token is revealed, then sampling a fixed-size mask without replacement with a softmax bias toward low-score downstream tokens. IM changes only mask selection; the host reward, mask count, diffusion-time schedule, and policy-gradient estimator remain unchanged.
来源：paper-fixed:PDF pp. 4–7, Sections 3.1–3.3, Algorithm 1, and experimental setup; PDF pp. 11–14, Appendix B
