StateBench: Testing how language models track state changes over long documents

Published 7 November 2026

The benchmark measures how models track entity state changes across long documents, finding frontier models achieve 13% or less on the hardest general-finite problems while gated-linear models perform near perfectly.