作者位次 #2/7ACL 2026Poster· Natural Language Processing
LLM Safety From Within: Detecting Harmful Content with Internal Representations
Difan Jiao · Yilun Liu · Ye Yuan · Zhenwei Tang · Linfeng Du · Haolun Wu · Ashton Anderson
摘要
Guard models are widely used to detect harmful content in user prompts and LLM responses. However, state-of-the-art guard models rely solely on terminal-layer representations and overlook the rich safety-relevant features distributed across internal layers. We present SIREN, a lightweight guard m...