SEAM
Measuring how much typed user speech gets absorbed into edited artifacts at unmarked within-turn seams.
Overview
When a user pastes a document and types an instruction in the same turn, the boundary between the two is unmarked. A model that misreads it edits the instruction into the artifact, or treats part of the artifact as a command.
Benchmark
SEAM measures how often that absorption happens and how far it spreads. The walkthrough and the model leaderboard are on the demo page.