Can self-report training help lie confession?

There is a growing body of work on introspective behavior in language models. I’m interested in whether self-report training, training a model to predict or explain its own behavior, can help it recognize and confess when it has lied.

My hypothesis is that confessing a lie requires the model to compare its previous answer either with its internal knowledge or with its stated principles, such as “I should never lie.” Self-report training may strengthen this kind of reflective process.

There are existing model organisms trained to lie or conceal secrets, which could provide a useful setting for testing this hypothesis.