MASCing: Configurable Safety Steering for Mixture-of-Experts Models via Activation Masks

arXiv·medium signalXBlueskyLinkedInCopy link

MoE architectures reduce inference cost through sparse activation but create new safety challenges — since only a subset of experts fire per token, traditional safety methods that target the full model can miss expert-specific failure modes. MASCing introduces activation steering masks that configure per-expert behavior without retraining, enabling granular safety controls on MoE models like Mixtral and Qwen-MoE. Practical for anyone deploying MoE models who needs fine-grained output control.