Tag
The paper identifies Harmfulness Propagation Dynamics in large language models and introduces Herald, a lightweight input moderator that uses cross-layer activation patterns to detect harmful prompts efficiently.