Brief

Researchers have developed new methods to steer refusal behavior in Mixture-of-Experts (MoE) large language models. They found that existing steering vector techniques remain effective even with MoE architectures. The proposed expert-aware methods further enhance this steering by leveraging specific expert routing patterns and directions, demonstrating that refusal signals can be effectively controlled by individual experts. AI

Researchers have developed RouteScan, a novel framework for auditing the safety of Mixture-of-Experts (MoE) Large Language Models (LLMs) without needing access to sensitive user data. This non-intrusive method analyzes low-level GPU execution telemetry, specifically the patterns of expert routing, to detect harmful behaviors. Evaluations on open-source MoE models show RouteScan achieves high generalization and accuracy, even on unseen harmful domains and novel jailbreak techniques, while offering a privacy advantage over content-based auditing. AI