Latent Knowledge

In exploratory LLMs fine-tuned to hide "secret words" (Taboo models), mechanical interpretation techniques like logit lens and SAE can read internal representations to extract the secrets quite effectively