Architecture Parameterization

Current Conference-Paper State

Source: path:paper/shared/figures/architecture-figure/architecture-autoregressive-compact.tex, input from path:paper/conference-paper/sections/03-methodology.tex.

Label: fig:architecture-ar-compact.

Caption headline: Architecture parameterization (compact autoregressive view).

The figure shows how the router and generator share one adapted Transformer backbone $\phi$ over the sequence $x_1,\ldots,\texttt{<lat>},E(z),\texttt{</lat>},y_1,\ldots,y_{t_y}$, with the router distribution $r_\phi(z|x)$ produced at the latent marker and the generator distribution produced autoregressively. It also shows the training-only posterior backbone $\xi$ that consumes $(x,y)$ and supplies $q_\xi(z|x,y)$.

Writing Role

This figure explains the implementation of the router-generator factorization and makes clear that the router and generator are roles of a shared adapted model rather than two independent model copies.

Stale Or Open Items

The figure caption uses $r_\phi$ and $q_\xi$ notation consistently with current notation decisions. The prose around it should reinforce that $p_\phi$ can denote the overall factorized model when the router/generator distinction is not the focus.