self-monitoring

Definition

Having a model watch its own outputs or internal states and flag errors, uncertainty, or unsafe trajectories as it goes. It is attractive where no external verifier exists, since the check costs only more of the model's own compute. The standing caveat is correlated failure: a model that is wrong for a reason it cannot represent is unlikely to catch itself, so self-monitoring supplements external verification rather than replacing it.

Explainers using this term

• Paper Walkthrough: The Personalization Mirage — LLMs Invent a Version of You, and Their Self-Reports Point the Wrong WayThe Personalization Mirage: How LLMs Fabricate User Profiles, and Why Self-Monitoring Misleads