Operationalizing “intent” to evade CoT monitors

Formalize and operationalize a measurable definition of a language model’s intent to evade chain-of-thought monitors that enables consistent empirical evaluation across models and setups.

Background

To study deliberate CoT control aimed at evading oversight, researchers need a clear, testable notion of “intent.” The authors note that existing concepts are not sufficiently concrete for measurement or comparison across systems.

A robust operationalization would support fair elicitation of evasion capabilities and clarify when a model is strategically suppressing or altering its CoT as opposed to failing randomly or following harmless instructions.