kNNGuard

Researchers introduced kNNGuard, a training-free guardrail for large language models that uses activation space to classify prompts, achieving up to 10x faster inference than fine-tuned methods. The s…