/p/2026-10-23 · explainer
Paper explainer · 2610.30026 · Zhao, Pandey and Misra

The highlighter was the whole signal.

An author marks the clauses that actually bind, in colour or with a highlighter. Then the file goes through a converter, the marks are gone, and the agent reads every sentence as equally important. Recovering those flagged requirements from 76 real policy documents scores 0.690 micro F1 when ingestion keeps how the text looked, 0.150 through a mainstream converter and 0.075 through plain text. Leave each cue out one at a time and only two of the six matter: highlight is worth 0.302 and colour 0.262, while bold, italic, monospace and strikethrough sit within 0.003 of nothing.

01 · The drop

choose the ingestion path

02 · Which cue carries it

tap a cue to drop it from ingestion

03 · Keep it, then name it

where it starts with the cue preserved and named

04 · How often it bites

05 · Your own stack illustrative

Results

What the paper actually measured

What it does not show

In practice