InnerExpert is introduced, the first method to leverage MoE-specific signals for per-token hallucination detection, and shows that InnerExpert outperforms existing methods across five datasets and two MoE architectures.
Abstract
Despite their widespread use, Large Language Models (LLMs) remain limited by a fundamental problem: the generation of plausible but false content, known as hallucinations. Most existing detection methods operate at the answer or sentence level, yet per-token detection is essential for localizing hallucinated spans and enabling fine-grained interventions. In this paper, we explore the use of the Mixture-of-Experts (MoE) paradigm to address this gap. In MoE architectures, a single forward pass activates a sparse subset of experts (i.e., distinct feedforward networks per layer) via a routing mechanism, producing internal signals (e.g., router entropy, expert disagreement, and expert usage patterns) that are unavailable in dense architectures and have not been previously exploited for hallucination detection. To this end, we introduce InnerExpert, the first method to leverage these MoE-specific signals for per-token hallucination detection. InnerExpert combines routing-level and standard transformer signals into compact per-token feature vectors, classified by a lightweight detector trained on labels produced by an LLM-as-a-judge pipeline, which enables continuous model updates without manual annotation. Our results show that InnerExpert outperforms existing methods across five datasets and two MoE architectures, achieving up to 0.91 answer-level and 0.76 token-level AUROC, while requiring only a single forward pass.
Object Hallucination in large vision-language models (LVLMs), where models generate non-factual content about input images, remains a critical barrier to their reliability in real-world applications. Existing mitigation strategies can be categorized into training-based and training-free methods. Training-based methods often achieve strong performance but are costly, requiring extensive computational resources, large-scale data, and time-consuming fine-tuning. Training-free approaches are particularly appealing due to their efficiency. However, existing training-free methods either require multiple decoding rounds, which adds computational overhead, or modify internal states in a model-specific way that risks degrading pretrained knowledge. We propose Test-Time Hallucination Mitigation (TTH) method, a novel training-free method that addresses both limitations. TTH introduces a token-validator module, implemented as a zero-shot Multi-Modal Classifier (MMC), to generate auxiliary logits grounded in the input image. These logits are fused with the original LVLM outputs at the token level for object tokens selected from a candidate pool. An entropy-based weighting scheme is then applied to enable robust and accurate predictions. Extensive experiments across multiple LVLM families and diverse benchmarks demonstrate that TTH consistently improves accuracy and robustness, underscoring its generalizability and practical effectiveness. Code is released at https://github.com/Mehran-TAM/TTH
Mehran Tamjidi, Hamidreza Dastmalchi, Ali Cheraghian et al.· 0 citations
LLMs have become more widespread in many critical areas, which has increased the risk of hallucinations, or instances where generated outputs contain false or misleading information. To address this problem, this paper introduces HalluciNet v3.0+, which is an adaptive multi-LLM hallucination detection framework that employs advanced preprocessing, feature engineering, and classification methods for strong factual verification. In the preprocessing stage, the system begins with normalizing raw LLM-generated outputs; using linguistic parsing through spaCy to break up the LLM outputs into structured claims; and then using Transformer-based tokenization for contextual refinement of the structured claims. For the feature extraction step, the system uses both BERT and BGE embeddings to create dense semantic representations of the structured claims and the context and relationship dependencies between the structured claims and the retrieved evidence. A hybrid feature selection is achieved by combining Recursive Feature Elimination (RFE) and attention-weight optimization methods, which provides for the selection of the most discriminative features while minimizing dimensionality. A novel ensemble-based classification solution is implemented using DeBERTa-v3 for Natural Language Inference, XGBoost for structured decision learning, and consensus scoring across multiple LLM sources to produce a final predicted output. A two-step RAG (Retrieval-Augmented Generation) module optimizes retrieval and retrieval based on the quality of the retrieved item in a cross-encoder-based framework while preserving the inherent factual quality of the item's contents. This system is designed to be fully scalable and production-ready and is therefore expected to provide trustworthy enhancements to the use of LLM-driven products.
S.Rajeshwari, M. K, Mohamed Habith S et al.· International Conference Com...· 0 citations
UniProbe is introduced, a lightweight, unified, learnable detector that models a frozen LVLM's heterogeneous computational trace from a single forward pass and achieves state-of-the-art token-level and object-hallucination detection.
D. Samuel, Guy Bar-Shalom, Fabrizio Frasca et al.· 0 citations
A lightweight linear detector is built on top of Role-Break that requires no fine-tuning of the VLM, whose feature dimension stays below 5,000 and reaches an average AUROC of 93.23 across six VLMs and four benchmarks.
Mingyu Wang, Weilin Jin, Wenbo Li et al.· 0 citations
Stability-Aware Adaptive Layer Aggregation (SALA) defines a layer-wise instability index computed from the training data, which measures how separability varies across domains and provides an upper bound on the possible separability drop from training to test.
Xinyi Li, Zhen Fang, Yadan Luo et al.· Proceedings of the 32nd ACM...· 0 citations
The D-Score is introduced, a simple spectral statistic computed from a single forward pass that is used as a hallucination score, classifying an input text as hallucinated when its D-Score is larger than a pre-defined quantity.
Bianca Raimondi, Davide Evangelista, Maurizio Gabbrielli et al.· 0 citations
