Decoding One Safety Trigger Token for Balancing Safety and Usability in Large Language Models
D-STT identifies and decodes specific 'safety trigger tokens'—the initial words in a model's refusal responses to harmful prompts—to activate its learned safety patterns during inference. By focusing on just one token, the method achieves robust defense against jailbreaks while maintaining high usability and low latency, making it a lightweight solution for proactive safety enforcement.