🤖 AI

Mixture-of-Expert Blocks Contain Strong Hallucination Detection Signals

This paper introduces InnerExpert, a novel per-token hallucination detection method that leverages unique internal signals from Mixture-of-Experts (MoE) architectures—such as router entropy and expert disagreement—to achieve superior performance across multiple datasets while requiring only a single forward pass.

Original authors: Joao Fonseca, Rodrigo Rodrigues, Paolo Romano

Published 2026-08-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Joao Fonseca, Rodrigo Rodrigues, Paolo Romano

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Large language models are the engines behind a new generation of artificial intelligence, capable of writing essays, solving problems, and holding conversations that feel remarkably human. Yet, beneath this fluency lies a persistent flaw: these machines sometimes invent facts with total confidence. This phenomenon, known as hallucination, occurs when a model generates content that sounds plausible but is entirely false. While these errors are often harmless in casual chat, they become dangerous when the model is used for medical advice, legal research, or news reporting. The core challenge for scientists is not just to spot these lies after the fact, but to catch them the moment they happen, ideally identifying the exact word where the truth begins to unravel.

For years, researchers have tried to solve this by asking the model to repeat its answer multiple times or by checking if different versions of the story agree with each other. These methods work, but they are slow and expensive, requiring the computer to do the same work several times over just to find a single mistake. A newer approach looks inside the model's own brain while it is thinking, searching for subtle signs of uncertainty in its internal calculations. However, most of these internal checks are limited to standard models. A different, more powerful type of architecture, called a Mixture-of-Experts, has recently become popular because it is faster and more efficient. This design works like a team of specialists, where a central manager decides which specific expert to call upon for each word being generated. Until now, the signals from this team dynamic were largely ignored for the purpose of catching lies.

In a new study, researchers from Portugal have developed a method called InnerExpert that finally puts these internal team signals to work. Instead of asking the model to repeat itself or waiting until the end of a sentence to check for errors, InnerExpert watches the model's internal routing process in real time. As the model generates a response, a small router decides which "expert" network should handle the current word. The researchers found that when the model is about to hallucinate, this router shows signs of confusion. The experts might disagree with one another, or the router might struggle to decide which specialist is best suited for the task. By tracking these moments of hesitation and disagreement, along with standard signals about how the model is focusing its attention, InnerExpert can assign a confidence score to every single word as it is created.

The team trained their system using a clever, automated approach. They fed the model thousands of questions and had a separate, highly capable artificial intelligence act as a judge to label which answers were true and which were made up. This allowed them to teach InnerExpert to recognize the specific patterns of internal signals that precede a hallucination, without needing any human to manually read and correct the data. Once trained, the system acts as a lightweight detector that runs alongside the main model, requiring only a single pass of data to produce a result. It does not need to slow down the model or ask it to generate extra text.

When tested on five different datasets and two different types of large models, InnerExpert proved to be remarkably effective. It successfully identified hallucinated answers with an accuracy score of up to 0.91 at the sentence level and 0.76 at the individual word level. These numbers represent a significant improvement over existing methods, which often struggle to pinpoint exactly where a lie begins. The system was able to generalize well, performing strongly even on questions about events that happened after the model was originally trained, a common scenario where hallucinations are most frequent. Crucially, it achieved this high level of detection while adding only a tiny amount of computational cost, making it practical for real-world use.

The study also revealed that no single signal from the model's internal team was enough to catch all the lies on its own. Some signals were better at detecting errors on one type of model, while others worked better on a different type. The true power of InnerExpert came from combining all these different signals—router confusion, expert disagreement, and usage patterns—into a single, unified score. This combination allowed the system to see a clearer picture of the model's confidence than any single metric could provide. The researchers demonstrated that by listening to the internal chatter of the model's experts, they could detect uncertainty with a precision that was previously impossible without slowing the system down.

This work suggests that the path to more reliable artificial intelligence may not require building larger models or more complex verification systems. Instead, it may lie in paying closer attention to the signals the models are already producing but ignoring. By learning to read the subtle signs of hesitation in a model's internal decision-making, we can build systems that know when they are guessing and when they are sure. The researchers note that while their method is highly effective, it is not a perfect solution, and future work will need to explore how well these techniques translate to different languages and more diverse tasks. For now, however, the study provides a clear and efficient way to catch hallucinations the moment they occur, offering a practical tool to make the outputs of artificial intelligence more trustworthy.

Technical Summary: InnerExpert – Mixture-of-Expert Blocks Contain Strong Hallucination Detection Signals

Problem Statement

Large Language Models (LLMs) frequently generate plausible but factually incorrect content, a phenomenon known as hallucination. While critical for reliability, detecting these hallucinations remains a challenge. Existing detection methods generally fall into three categories:

Sampling-based: Methods like SelfCheckGPT and Semantic Uncertainty probe consistency across multiple stochastic generations. While effective, they are computationally expensive at inference time due to the requirement of multiple forward passes.
Internal-signal: Methods like LLM-Check and MIND operate on hidden states and attention patterns within a single forward pass. However, they typically utilize a limited subset of signals available in dense architectures and ignore the unique routing structures of Mixture-of-Experts (MoE) models.
Trainable detectors: Approaches like HaluNet and FacLens learn classifiers over dense transformer signals but often operate at the answer or sentence level rather than the token level, limiting their ability to localize specific hallucinated spans.

Furthermore, most current methods lack per-token granularity, which is essential for fine-grained localization of hallucinated spans and enabling targeted interventions. There is a specific gap in leveraging the internal signals unique to MoE architectures for this purpose.

Methodology: InnerExpert

The paper introduces InnerExpert, a single-pass, per-token hallucination detector designed specifically for MoE architectures. It operates by intercepting the model's forward pass to extract and combine internal signals into a compact feature vector, which is then classified by a lightweight detector.

1. Signal Extraction

InnerExpert distinguishes between standard signals (available in any transformer) and MoE-specific signals (arising from the routing mechanism).

Standard Signals:
- Hidden State Score: Captures representational redundancy by computing the mean log-singular-value of the centered covariance matrix of hidden states up to the current token.
- Attention Score: Captures focus patterns via the log-cumulative sum of the diagonal entries of the attention kernel.
MoE-Specific Signals: These signals are unavailable in dense architectures and are central to the paper's contribution. They include:
- Router Entropy: The entropy of the gating distribution, indicating uncertainty in expert selection.
- Expert Hidden Score: A routing-weighted aggregation of hidden state scores computed independently for each activated expert.
- Expert Similarity: A routing-weighted average of pairwise cosine similarities between activated experts' outputs; low similarity indicates disagreement (analogous to ensemble disagreement).
- Expert Usage Distribution: A cumulative distribution of expert selections weighted by routing probabilities.
- Gini Impurity & Inverse Herfindahl Index: Metrics derived from the usage distribution to measure the spread and effective number of contributing experts.

2. Feature Assembly and Classification

For each generated token, the per-layer signals are flattened into a single feature vector $\Phi_t$ . This vector is fed into a lightweight classifier (e.g., Logistic Regression, XGBoost, MLP) trained to output a hallucination score $s(y_t) \in [0, 1]$ . The answer-level score is obtained by aggregating token scores (using the max operator).

3. Training Supervision

To avoid manual annotation, the authors employ an unsupervised training approach using an "LLM-as-a-judge" pipeline:

Data Source: Questions from the RealtimeQA platform (Jan 2024–Dec 2025).
Generation: The host model generates answers both with and without reference evidence (simulating RAG).
Labeling: A judge LLM (GLM-5.1) compares the generated answer against reference evidence to produce binary hallucination labels and identify hallucinated spans. These spans are mapped to per-token binary labels.
Validation: The quality of these labels is validated against human annotations, showing significantly higher agreement than weak lexical/semantic heuristics.

Key Contributions

InnerExpert Framework: The first method to leverage MoE-specific internal signals (router entropy, expert disagreement, usage patterns) for per-token hallucination detection without modifying the host LLM or requiring additional sampling.
Unsupervised Training Pipeline: A formalized approach for training per-token detectors using LLM-as-a-judge labeling against reference evidence, eliminating the need for manual annotation.
Signal Inventory and Analysis: The formalization of six MoE-specific signals alongside standard transformer signals, enabling a systematic study of which routing signals carry discriminative information for hallucinations.
State-of-the-Art Performance: Demonstration that combining these signals yields superior performance compared to baselines across multiple datasets and architectures.

Experimental Results

The authors evaluated InnerExpert on two open-weight MoE models (OLMoE-1B-7B and Gemma-4-26B) across five datasets (RealtimeQA, SQuAD, TruthfulQA, NQ-Open, FreshQA).

Performance: InnerExpert achieved up to 0.91 answer-level and 0.76 token-level AUROC on average across five datasets. Specifically, the best-performing variant (IE XGB) achieved an average answer-level AUROC of 0.912 on Gemma-4-26B, while the best variant for OLMoE-1B-7B (IE LR) achieved an average token-level AUROC of 0.762. It consistently outperformed baselines spanning sampling-based, internal-signal, and trainable paradigms.
Signal Contribution: No single MoE signal was uniformly strong across both architectures. However, combining them via the classifier yielded consistent gains, indicating that the signals provide complementary information. For instance, on OLMoE, the combination improved AUROC by 0.197 over the best individual MoE signal.
Efficiency: InnerExpert requires only a single forward pass. While extracting full MoE signals increases inference cost to approximately $2.5\times$ vanilla generation, the total overhead (including classifier inference) remains around $3\times$ vanilla. This is significantly more efficient than sampling-based methods, which require multiple forward passes (e.g., $K=5$ ) and often double GPU memory usage.
Generalization: The method maintained consistent advantages over baselines on temporally out-of-distribution data (post-training events) and cross-dataset generalization tasks.

Significance and Claims

The paper claims that InnerExpert addresses the critical gap in hallucination detection by exploiting the unique internal structure of MoE models. By utilizing signals related to epistemic uncertainty (specifically expert disagreement and routing uncertainty), the method achieves fine-grained, per-token detection with high efficiency.

The authors emphasize that:

MoE-specific signals are a rich, previously unexplored resource for detecting hallucinations in text generation.
A simple combination of these signals with standard transformer signals via a lightweight classifier is sufficient to achieve state-of-the-art results, without needing complex architectures like HaluNet.
The unsupervised labeling pipeline allows for continuous model updates without manual effort, making the approach scalable.

The work is modest in its scope, noting limitations such as the small scale of human validation, restriction to English, and the reliance on lightweight classifiers. The authors specifically highlight that MLP and Logistic Regression (LR) are the most consistent performers across models, whereas XGBoost suffers from sharp F1 degradation on certain models (e.g., Gemma), indicating threshold sensitivity rather than a lack of discriminative power. The authors suggest future work could focus on domain certification, controlled text generation, and expanding to multilingual settings.