RASET: Router-Agnostic Safety-Critical Expert Tuning Exposes Localized Safety Enforcement Failures in Mixture-of-Experts LLMs
The episode analyzes RASET, a method that examines Mixture-of-Experts LLMs. Researchers found that AI safety mechanisms are not controlled by a single refusal expert or router path. Instead, safety is localized within the topic-specialized experts themselves. RASET proposes surgically tuning these specific experts to improve security.