Likelihood Ranking doesn't Scale Like Prompting in LLMs
It is suggested that likelihood preferences over controlled declarative alternatives and task-conditioned answer selection probe distinct aspects of model behavior, and should not be treated as interchangeable.