Lack of Introspection: Unless specifically instrumented, transformer-based LLMs have no ability to explicitly access their own internal states—the activations in their feed-forward layers, attention ...