Safety-Oriented Routing Analysis of Mixtral MoE Under Benign and Harmful Prompts
Researchers have analyzed the routing behavior of the Mixtral 8x7B-Instruct model when presented with both benign and harmful prompts. They used activation-based and gradient-based signals to understand how the model selects experts for processing different types of input. The study found that while most experts are shared between benign and harmful prompts, a small subset shows distinct preferences. Interventions to suppress these preferred experts reduced harmful responses, indicating that safety-relevant routing is subtle and distributed across layers. AI
IMPACT Provides insights into the internal workings of Mixture-of-Experts models, potentially informing future safety research and development.