@GuardReasoner

Researchers have developed COLAGUARD, a new safety guardrail for large language models that transfers multi-step reasoning into a continuous latent space to reduce latency. In tests across ten moderat…