Part 2 · Intuition
Head-to-head
| Autoregressive (typewriter) | Masked diffusion (editor) | |
|---|---|---|
| Generation order | Strictly left → right | Any order; easiest-first, coarse-to-fine |
| Forward passes for N tokens | N, strictly sequential | k steps with k ≪ N; k is tunable |
| Context per prediction | Past tokens only | Both directions |
| Revising committed text | Impossible | Yes, via remasking |
| Speed–quality trade-off | Fixed by length | Dial: number of steps |
| Variable length + KV caching | Natural | Needs Block Diffusion (see the papers) |
Honesty box. Autoregressive models still hold the overall quality crown at the frontier, and vanilla masked diffusion arrives with real handicaps: it generates fixed-length canvases, it can’t reuse cached computation the way AR decoding does, and its likelihoods historically lagged. The reason Inception’s paper list is interesting is that nearly every entry knocks down one of these handicaps.
That is the intuition in full. The next two sections re-tell the same story with the equations attached: first the continuous case, where diffusion and its vocabulary were invented, then the discrete, masked case that a dLLM actually runs. They are the most demanding sections of this page — and the most rewarding, because after them the nine papers read like straightforward engineering decisions rather than magic.