1 paper · 1 filter
Jeffrey Olmo, David Zhang, Tyler Tracy
Control evaluations measure whether safeguards can prevent an AI model from causing undesired outcomes even when it is intentionally trying to do so. These evaluations typically us…