hallucination detection

Foundation models frequently exhibit “hallucinations” during autonomous decision-making, leading to high-risk misjudgments—yet no formal definition or systematic characterization of hallucination exists for decision tasks. Method: This work introduces the first task-specific definition of hallucination in decision-making and establishes a cross-task, scalable hallucination taxonomy; proposes a synergistic framework integrating uncertainty quantification with hallucination detection; and develops a joint detection–decision evaluation paradigm grounded in systematic survey analysis, probabilistic uncertainty modeling, and decision-chain interpretability. Results: We present the first comprehensive landscape of hallucination detection techniques tailored to decision contexts, contributing seven actionable implementation guidelines and identifying five critical research gaps—thereby advancing the safe deployment of trustworthy foundation models in high-stakes domains such as healthcare and transportation.

Hallucination—generating factually incorrect or fabricated content that appears plausible—severely undermines the reliability and trustworthiness of large language models (LLMs). Method: This work establishes the first universal theoretical framework for LLM hallucination, formally defining its essence and proving its intrinsic inevitability within computable models. It innovatively distinguishes *intrinsic* from *extrinsic* hallucination and rigorously clarifies the conceptual boundaries between *factual accuracy* and *faithfulness*. A fine-grained taxonomy is developed, covering cross-modal and multi-task scenarios. The analysis integrates theoretical modeling, classification-based formalization, and empirical validation—including data provenance tracing, logical consistency checking, benchmark evaluation, and human-subject experiments. Contribution/Results: The study systematically uncovers root causes and human perception mechanisms of hallucination. It releases an open-source evaluation benchmark and an online resource platform, providing a unified theoretical foundation and reusable toolset for hallucination detection, mitigation, and governance.

This paper addresses the challenge of hallucination detection in large language model (LLM) inference. Methodologically, it proposes a lightweight, unsupervised automatic detection framework: first, a systematic hallucination taxonomy is constructed, and diverse hallucinated responses are controllably generated via prompt engineering; second, response embeddings—obtained using a pre-trained embedding model—are projected into a low-dimensional space, where dimensionality reduction and unsupervised clustering reveal distributional separability between hallucinated and truthful responses; finally, the distance from each response’s embedding to the centroid of truthful responses quantifies its hallucination severity. Key contributions include: (i) the first demonstration of controllable hallucination re-generation via prompt engineering; (ii) empirical discovery of a strong correlation between hallucination severity and embedding-space distance to the truthful centroid; and (iii) validation that even simple classifiers achieve high detection accuracy, confirming the framework’s effectiveness in enhancing LLM reliability and its practical deployability.

Deep learning models in medical AI devices frequently generate clinically harmful “hallucinations”—outputs that appear plausible but mislead clinical decision-making—yet these phenomena lack a rigorous, standardized definition and systematic evaluation. Method: We propose the first unified, cross-modal and cross-task definition of hallucination in medical devices: data-driven model outputs exhibiting superficial plausibility yet potentially compromising clinical judgment; we further distinguish *impactful* from *harmless* hallucinations. Through dual-track empirical analysis—spanning imaging tasks (e.g., lesion segmentation) and non-imaging tasks (e.g., physiological parameter prediction)—we develop an integrated theoretical–empirical assessment framework. Contribution/Results: We introduce the first systematic, multi-product-line hallucination taxonomy and quantitative evaluation paradigm; characterize the hallucination spectrum across diverse clinical scenarios; and deliver actionable detection protocols and mitigation strategies—providing a methodological foundation for regulatory review and safety governance of medical AI.

This work addresses the critical challenge of hallucination in large language models during tool-augmented reasoning, which often leads to incorrect tool selection, erroneous parameter specification, or unjustified tool bypassing—thereby compromising system reliability. The authors propose a novel real-time hallucination detection method that leverages internal model representations from a single forward pass, eliminating the need for additional inference steps or external validation. By deploying a lightweight classifier to analyze activation patterns within the same inference cycle, the approach efficiently identifies hallucinations at both the tool-selection and parameter-specification levels. Evaluated across diverse domains, the method achieves up to 86.4% detection accuracy, substantially outperforming existing techniques while introducing negligible inference latency, thus significantly enhancing both the safety and efficiency of deployed systems.

Existing geometric hallucination detection metrics struggle to distinguish specific hallucination types in the absence of ground truth and are highly sensitive to domain shifts. This work addresses these limitations by constructing a synthetic dataset to systematically evaluate the capacity of various geometric statistics to capture key hallucination attributes—such as output correctness, relevance, and coherence—and reveals that different metrics align with distinct hallucination types. Furthermore, the study proposes a simple yet effective normalization strategy that substantially mitigates the impact of domain shift. Experimental results demonstrate that, under multi-domain settings, the proposed approach improves AUROC by 34 percentage points, significantly enhancing the cross-domain robustness of geometric hallucination detection metrics.

Visual hallucinations—perceptual experiences of identifiable stimuli that lack external sensory input—are strongly associated with multiple neuropsychiatric disorders, yet their underlying neural mechanisms remain poorly understood. To address this, we propose the first neurobiologically inspired computational model that formally attributes human visual hallucinations to endogenous adversarial interactions among brain regions, operationalized as a generative adversarial network (GAN). Integrating principles from computational neuroscience, dynamic functional brain region modeling, and GAN architecture, our model successfully reproduces hallmark hallucination dynamics: spontaneous emergence, morphological evolution, and stimulus dependence. This work advances a novel mechanistic hypothesis for hallucination-related psychiatric conditions and establishes a new paradigm for developing neuroscientifically interpretable artificial intelligence models.

In enterprise settings, large language models (LLMs) suffer from hallucination due to limited context windows and outdated knowledge; existing mitigation strategies—such as gold-standard QA repositories or secondary verification models—are costly and lack formal guarantees of correctness. This paper proposes an interactive, visualization-enabled knowledge graph framework for hallucination detection: LLM-generated assertions are dynamically linked to proprietary knowledge sources to construct a structured truth-graph, supporting confidence scoring, provenance tracing, and human-in-the-loop feedback. Our key contributions lie in the integration of adaptive knowledge graph construction, interpretable natural language understanding (NLU), and human–AI collaborative diagnosis—enabling real-time identification and auditable verification of hallucinated content. Experiments demonstrate significant improvements in LLM response trustworthiness and reliability under constrained context and knowledge inconsistency, while establishing a sustainable, feedback-driven optimization loop.

Existing hallucination detection methods focus narrowly on factual consistency, overlooking potential creative value and struggling to balance accuracy with creativity across diverse scientific tasks. Method: We propose HIC-Bench—a novel benchmark framework that systematically distinguishes *Intelligent Hallucination* (IH) from *Defective Hallucination* (DH). It evaluates both dimensions across ten open-ended scientific innovation tasks using a dual-axis metric: creativity (integrating Torrance Tests of Creative Thinking with hallucination-specific dimensions) and factual deviation. Innovations include an IH/DH binary classification paradigm, Dynamic Hallucination Prompting (DHP), a multidimensional metrics matrix, cross-disciplinary task design, ensemble evaluation by multiple LLMs, and human validation. Results: IH and DH exhibit a nonlinear relationship; creativity and factual accuracy can be jointly optimized; and hallucinations—when appropriately structured—can serve as catalysts for scientific innovation.

This work addresses the challenge of hallucinations—factual inaccuracies in text generated by large language models—and proposes CHAI R, a novel decoding-time detection framework. CHAI R is the first to systematically leverage statistical features derived from multi-layer internal logits, such as extrema, mean, standard deviation, and slope, to construct compact representations for hallucination identification via supervised learning. Evaluated on TruthfulQA and MMLU benchmarks, the method significantly improves detection accuracy and demonstrates strong generalization capabilities, particularly in zero-shot settings. By exploiting internal model representations, this approach offers a promising direction for adaptive decoding strategies aimed at enhancing the factual consistency of language model outputs.

This work addresses the critical reliability concerns of generative MRI reconstruction models, which are prone to producing hallucinated anatomical structures under minute input perturbations—posing significant risks of clinical misdiagnosis. For the first time, the study systematically employs adversarial perturbations to actively induce and quantify such hallucinations, revealing the models’ extreme sensitivity to subtle input changes. Experiments on the fastMRI dataset, conducted on both UNet and end-to-end VarNet architectures using adversarial example generation techniques, demonstrate that current models frequently generate clinically unreliable hallucinatory content. Notably, conventional image quality metrics fail to effectively detect these errors. This research establishes a novel paradigm for evaluating and enhancing the trustworthiness of medical image reconstruction systems.

This paper identifies a persistent hallucination phenomenon in large language models (LLMs) induced by spurious correlations in training data—e.g., superficial statistical associations between surnames and nationalities—leading models to generate incorrect answers with high confidence. Critically, this behavior is scale-invariant, resistant to standard confidence-based filtering and internal state probing, and persists even after refusal fine-tuning. Method: We introduce a novel methodology combining synthetic controlled experiments with theoretical analysis to systematically evaluate mainstream open- and closed-source models—including GPT-5—under spurious correlation scenarios. Contribution/Results: Our analysis reveals, for the first time, how spurious correlations mechanistically undermine the reliability of both model confidence and internal representations. The hallucinations exhibit strong robustness and cross-model generality. Consequently, we argue that hallucination detection must shift toward statistically aware paradigms explicitly designed to mitigate bias-induced failures.