Investigating task-specific prompts and sparse autoencoders for activation monitoring

Henk Tillman, Dan Mossing

Language models can behave in unexpected and unsafe ways, and so it is valuable to monitor their outputs. Internal activations of language models encode additional information that…