Language models (the engines behind ChatGPT) struggle to keep a coherent mental notepad over long texts. This is state tracking: the ability to manage changing variables, like following a recipe or running code.
Researchers created a four-level hierarchy for these tasks, from simple constant registers to fully general variable swapping. They built a , StateBench, and tested models from open-source to frontier giants.
No tested model reliably handled self-referential assignments (where a value depends on its previous state) over long sequences. This points to a need for explicit external memory or tools for reliable state management.