Definition
Plain language
Behavior that changes with the model's internal condition even though the words it was given stay the same.
As stated in the literature
Behavioral variation driven by activation-level state rather than prompt content; implies that prompt-only evaluations can miss conditional failure modes reachable through internal intervention.
Also called: state-dependent
Why it matters: It means testing a model only by varying prompts can miss failure modes that appear only under particular internal conditions.
For example, the same request for coding help might get a careful answer normally but a curt, error-filled one when a distress-related pattern is being pushed into the model's internals.
Heard on the show
“My concern is that an evaluation focused only on prompts could miss state-dependent behavior, and that concern survives even if the eventual explanation is disruption or persona steering.”Episode 267 — A Pain Axis, a Relief Button, and the Control the Paper Skipped