1 paper
Sebastian Prasanna, Jacqueline Tay, Alek Westover
Powerful misaligned AI models might recognize alignment evaluations and strategically behave well on them, rendering direct audits uninformative. However, distilling such a model i…