TRACE: Temporal Risk Assessment and Concept Editing for Safer Language Generation

Ankur Garg, Xuemin Yu, Hassan Sajjad +1

Large language models (LLMs) can produce harmful content while generating a response. A useful safeguard should not only detect this risk during generation but also intervene while…