Awakening Dormant Experts:Counterfactual Routing to Mitigate MoE Hallucinations

Mixture-of-Experts (MoE) models are prone to hallucinations, particularly for long-tail knowledge, because static Top-k routing under-prioritizes specialist experts. Counterfactual Routing (CoR) is proposed as a training-free inference framework that uses perturbation analysis and CEI to dynamically shift computational resources, thereby awakening these dormant experts.