This research introduces a novel approach for LLMs to self-assess confidence and intelligently route requests to appropriate experts when uncertain.

• Enables AI systems to recognize their own limitations and redirect questions when necessary
• Implements special confidence tokens that help models express uncertainty
• Creates more reliable AI systems for high-stakes applications
• Demonstrates significant improvements in identifying untrustworthy outputs

For security applications, this approach is crucial as it provides a mechanism to prevent potentially harmful or incorrect AI responses in critical situations by defaulting to safer alternatives when confidence is low.

Learning to Route LLMs with Confidence Tokens