Value Probes preprint reports roughly 50% higher peak AI audit scores

A preprint reports that a method called Value Probes raises peak language-model audit scores by roughly 50% against standard approaches using matched token budgets, across 18 target behaviors and 3 model scales. The method searches multiple conversation paths and uses signals from the target model's internal activations to estimate which paths could lead to problematic behavior. The researchers also report that probes can be trained with data from smaller models. The measured improvement concerns peak audit scores, rather than a demonstrated reduction in harmful behavior after deployment.