Routing-Aware Safety Alignment for Mixture-of-Experts Models
Mixture-of-Experts (MoE) models face safety issues because standard fine-tuning can create 'alignment shortcuts' where safety improves through routing changes, not by fixing unsafe experts. RASA is a framework that fixes this by selectively fine-tuning only the most problematic experts and enforcing consistent routing, leading to robust safety against diverse jailbreaks.