Diffusion Language Model Unmasking-Order Distillation Diagram
Overview
A diagram of distilling a faster unmasking schedule into a masked diffusion language model, on the example answer "So 12 − 2 = 10 eggs left". It is an editable PowerPoint diagram adapted from the original paper figure under CC BY 4.0.
What the diagram shows
Left: an unmasking trajectory s0 → s1 → s2 → s3 of token rows (grey shade = surprisal, teal = revealed, orange = promoted) and a natural-vs-promoted round table (6 rounds → 3 rounds). Upper right, single-step promotion: candidate positions 𝒞r with pθ0(yi | sr) ≥ pmin and rank ≤ k under a frozen teacher, two trial reveals with before / after other-position surprisal bars (smaller drop vs. largest positive drop, selected), and the rule i★r ∈ argmaxi∈𝒞r Σj∈ℳr∖{i} [hj(sr) − hj(s(i)r)]. Lower right: the student pθ distilled from trajectory states, a parallel decoder, a training box with ℒ = ℒ+ + βℒ− (predict masked tokens on promoted trajectories, KL to the frozen teacher on failed ones) and an inference box that commits position j when maxv pθ(v | st, j) ≥ τ in one forward pass. Every box, arrow, label and math term is a native shape you can recolor, resize and relabel for your own slides, posters or papers.
File format
The download contains .pptx (1 slide, 13.33 × 9.01 in; opens in PowerPoint, Keynote, WPS and LibreOffice Impress; Google Slides works too, but complex freeforms may shift slightly), .svg and .drawio. Includes preview.png (full resolution) and a 1200 px web preview.
Shapes and editing
The file has 294 native editable shapes: 159 auto-shapes, 74 text boxes, 15 freeforms, 46 lines/connectors. No embedded images. Everything is vector. You can edit all text and math (s0 … s3, 𝒞r, pθ0(yi | sr) ≥ pmin, the stacked argmax / Σ selection rule with s(i)r, ℒ = ℒ+ + βℒ−, the commit rule with τ), every token tile and round-table cell, the surprisal bar charts, dashed candidate outlines, teacher / student model stacks with native vector flame and snowflake icons, the threshold bars, curved trajectory arrows, dashed distillation routes and the legend.
Embedded images and fonts
This file has no embedded images; everything is vector. Body text uses Calibri and formulas use Times New Roman. Calibri ships with Microsoft Office and Times New Roman is preinstalled on Windows and macOS (LibreOffice substitutes the metric-compatible Carlito and Liberation Serif). The calligraphic 𝒞, ℳ and ℒ use Cambria Math, which ships with Microsoft Office; other apps substitute a math font. Sub- and superscripts are real formatted text runs, not equation objects. The SVG keeps text as text; the draw.io file is a shape-by-shape conversion and may need small nudges. This is a faithful redraw for reuse, not the authors' original source file.
Source & license Adapted from Figure 3 of Learning When to Commit: Schedule Distillation for Fast Diffusion Language Models by Aidar Myrzakhan, Zhiqiang Shen (arXiv:2610.35804, https://arxiv.org/abs/2610.35804). Licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Redrawn as editable shapes; changes made by 叫我Alonzo就好了/alonzolab. Not endorsed by the original authors.
Questions or feedback? Email [email protected]