Token-Level Surprisal-Share Weighting

Current Conference-Paper State

Source: path:paper/shared/figures/token-weighting/figure.tex, input from path:paper/conference-paper/sections/03-methodology.tex.

Label: fig:token-weighting.

Caption headline: Token-level surprisal-share weighting concentrates reconstruction pressure on branching tokens.

The figure uses a tree panel and token strip to show how base-model surprisal $b_t$ and share weights $w_t \propto b_t$ focus reconstruction pressure on positions where $z$ has room to explain $y$ beyond $x$.

Writing Role

This is the methodology figure for token-level $p_\theta$ surprisal weighting. It should be coordinated with the objective language in Base-directed reconstruction objective and the method-variant language in Base-directed reconstruction variants.

Stale Or Open Items

The current caption uses $p_{\text{base}}$ notation. Current notation decisions prefer $p_\theta$ for the fixed base model, so the figure caption and labels should be checked for notation consistency.