MoE Models Offer Strong Signals for Hallucination Detection

Hallucinations remain a significant challenge for Large Language Models (LLMs), where plausible but false content is generated. While most detection methods operate at the answer or sentence level, this research focuses on the crucial need for per-token detection to precisely locate and address hallucinated spans. The paper explores the unique advantages of Mixture-of-Experts (MoE) architectures in this context. MoE models, by design, activate only a sparse subset of experts during a forward pass, generating internal signals such as router entropy, expert disagreement, and usage patterns that are unavailable in dense Transformer architectures. The proposed method, InnerExpert, is the first to exploit these MoE-specific signals for per-token hallucination detection. InnerExpert combines these routing-level signals with standard Transformer signals into compact feature vectors, which are then classified by a lightweight detector. This detector is trained using an LLM-as-a-judge pipeline, allowing for continuous model updates without manual annotation. Results show InnerExpert significantly outperforms existing methods across various datasets and MoE architectures, achieving high AUROC scores for both answer-level and token-level detection with minimal computational overhead.