Researchers have developed a new method called InnerExpert that can detect hallucinations in large language models (LLMs) at the level of individual tokens. Hallucinations occur when LLMs generate plausible but false content. The method uses signals from Mixture-of-Experts (MoE) architectures, which are not available in traditional dense architectures. These signals include router entropy, expert disagreement, and expert usage patterns. InnerExpert combines these signals with standard transformer signals to create compact feature vectors that can be classified by a lightweight detector. This allows for continuous model updates without manual annotation. --- Why it matters: This matters because detecting hallucinations at the token level is essential for fine-grained interventions in LLMs. Current detection methods often operate at the answer or sentence level, which can make it difficult to localize and correct hallucinated spans. Source: https://arxiv.org/abs/2608.17687