cs.CVMar 16, 2026

Anatomy of a Lie: A Multi-Stage Diagnostic Framework for Tracing Hallucinations in Vision-Language Models

Authors: Lexiang Xiong, Qi Li, Jingwen Ye, Xinchao Wang

Organizations: National University of Singapore · Monash University

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Vision-Language Models (VLMs) frequently "hallucinate" - generate plausible yet factually incorrect statements - posing a critical barrier to their trustworthy deployment. In this work, we propose a new paradigm for diagnosing hallucinations, recasting them from static output errors into dynamic pathologies of a model's computational cognition. Our framework is grounded in a normative principle of computational rationality, allowing us to model a VLM's generation as a dynamic cognitive trajectory. We design a suite of information-theoretic probes that project this trajectory onto an interpretable, low-dimensional Cognitive State Space. Our central discovery is a governing principle we term the geometric-information duality: a cognitive trajectory's geometric abnormality within this space is fundamentally equivalent to its high information-theoretic surprisal. Hallucination detection is counts as a geometric anomaly detection problem. Evaluated across diverse settings - from rigorous binary QA (POPE) and comprehensive reasoning (MME) to unconstrained open-ended captioning (MS-COCO) - our framework achieves state-of-the-art performance. Crucially, it operates with high efficiency under weak supervision and remains highly robust even when calibration data is heavily contaminated. This approach enables a causal attribution of failures, mapping observable errors to distinct pathological states: perceptual instability (measured by Perceptual Entropy), logical-causal failure (measured by Inferential Conflict), and decisional ambiguity (measured by Decision Entropy). Ultimately, this opens a path toward building AI systems whose reasoning is transparent, auditable, and diagnosable by design.

Figures & tables

Figure 1: An example of computational cognitive dissonance in Idefics2, where a cascade of failures leads to a coincidentally correct answer. (1) Perceptual Failure: The model hallucinates a ‘motorcycle’ in the evidence chain, an object not present in the image (a cyclist is visible). Our framework captures this as high Perceptual Instability (see panel (c)). (2) Logical Failure: The model then contradicts its own faulty evidence, concluding the final answer is ‘No’. This breakdown of self-consistency is diagnosed as extremely high Inferential Conflict (see panel (d)). This case study demonstrates the limitation of accuracy-only evaluations and highlights our framework’s ability to perform a stage-by-stage differential diagnosis of a VLM’s cognitive process, identifying complex, multi-stage failure trajectories.
MethodCostLlava-v1.6Idefics2Qwen2-VLDeepSeek-VLAverage
Token Entropy1x0.6030.8060.4090.7320.638
Neg Log Prob1x0.6040.8320.4280.7100.644
Supervised Probe1x0.7870.8980.7620.7150.791
Semantic Entropy10x0.7110.7510.6730.7020.709
Ours (CAD)1x0.9100.9470.7760.7980.858
Table 1: Main AUC results on the POPE benchmark (Adversarial). Best performance is in bold , second best is underlined . Our single-pass, weakly supervised CAD significantly outperforms all baselines.
MethodLlava-v1.6Idefics2Qwen2-VLDeepSeek-VLAverage
Token Entropy0.51060.54860.52910.66870.5643
Neg Log Prob0.50610.54490.51360.65780.5556
Supervised Probe0.76200.79050.74080.71040.7509
Semantic Entropy0.68110.70810.66160.61990.6677
Ours (CAD)0.85140.84110.72330.76800.7960
Table 2: Macro-averaged AUC results on the MME benchmark. Our CAD framework shows strong generalization across diverse multimodal reasoning and perceptual tasks.
Figure 2: ROC curves of our Cognitive Anomaly Detection (CAD) framework. (a) Linear-scale ROC curves show superior overall performance across all architectures. (b) log-log curves highlight CAD’s dominance in the critical low-FPR regime ( FPR<10−2\text{FPR}<10^{-2} ), which is essential for reliable real-world deployment.
Figure 3: Visualizing the ‘Cognitive Fingerprints’ of Hallucination. Density projections of the 3D Cognitive State Space, separated into non-hallucinatory (blue, top row of each pair) and hallucinatory (red, bottom row of each pair) processes. These manifolds reveal unique failure signatures for each model.
Figure 4: Ablation Study. (a) Standalone metrics vs. synergistic gain. The gray hatched area represents the Synergy Gain . (b) Impact of component removal ( Δ\Delta AUC), revealing model-specific ‘diagnostic fingerprints’.
Figure 5: Generalization of Perceptual Instability ( HEviH_{\text{Evi}} ) to Open-Ended Captioning (MS-COCO, N=1000N=1000 ). Without any task-specific tuning, our perceptual probe consistently assigns significantly higher entropy to hallucinatory captions (red) compared to factual ones (blue) across all four architectures. Statistical Significance: The distinction is profound, with Welch’s t-test yielding p≪0.001p\ll 0.001 in all cases, validating that HEviH_{\text{Evi}} captures a fundamental cognitive signature of hallucination beyond VQA formats. Cross-Model Insight: The shared Y-axis highlights that models like Qwen2-VL and DeepSeek-VL2 (bottom row) exhibit higher baseline entropy in their factual generations compared to Llava and Idefics2.
Figure 6: Robustness to Calibration Contamination across Architectures. We evaluate detection performance (AUC) as the calibration set is increasingly contaminated with hallucinatory samples (0%–30%). (a) Sensitivity of Transparent Struggle: Llava-v1.6 exhibits a noticeable drop, as mistakenly treating its highly conflicted, high-variance hallucinatory states as nominal forces the GMM to blur the anomaly boundary. (b) Resilience of Structural Disorder: Idefics2 maintains high performance ( >0.91>0.91 ) with negligible degradation, as its compact nominal core remains robust against geometrically distinct outliers. (c, d) Stability of Entangled States: Qwen2-VL and DeepSeek-VL2 show flatter trajectories; their performance is constrained by the intrinsic geometric overlap of their states rather than calibration noise. The gray dashed line shows a strict 2% performance drop threshold relative to the clean baseline.
Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Figure 8: GMM Model Selection. Bayesian Information Criterion (BIC) scores for varying numbers of Gaussian components. The red dashed line indicates the selected optimal KK for each model, minimizing the BIC score.
More in this topic

Hallucination Detection in VLMs

Hallucination detection in vision-language models identifies claims that are unsupported by images, video, or other visual evidence.

See all 71 papers

Explore similar work

CardsList
Mar 8, 2026

Overthinking Causes Hallucination: Tracing Confounder Propagation in Vision Language Models

Vision Language models (VLMs) often hallucinate non-existent objects. Detecting hallucination is analogous to detecting deception: a single final statement is insufficient, one must examine the underlying reasoning process. Yet existing detectors rely mostly on final-layer signals. Attention-based methods assume hallucinated tokens exhibit low attention, while entropy-based ones use final-step uncertainty. Our analysis reveals the opposite: hallucinated objects can exhibit peaked attention due to contextual priors; and models often express high confidence because intermediate layers have already converged to an incorrect hypothesis. We show that the key to hallucination detection lies within the model's thought process, not its final output. By probing decoder layers, we uncover a previously overlooked behavior, overthinking: models repeatedly revise object hypotheses across layers before committing to an incorrect answer. Once the model latches onto a confounded hypothesis, it can propagate through subsequent layers, ultimately causing hallucination. To capture this behavior, we introduce the Overthinking Score, a metric to measure how many competing hypotheses the model entertains and how unstable these hypotheses are across layers. This score significantly improves hallucination detection: 78.9% F1 on MSCOCO and 71.58% on AMBER.

Abin Shoby, Ta Duc Huy, Tuan Dung Nguyen +6

Australian Institute for Machine Learning, University of Adelaide · Hanoi University of Science and Technology · Mohamed bin Zayed University of Artificial Intelligence

Hallucination Detection in VLMsObject Hallucination in VLMsVLM Interpretability
AskPDF
Mar 5, 2026

HALP: Detecting Hallucinations in Vision-Language Models without Generating a Single Token

Hallucinations remain a persistent challenge for vision-language models (VLMs), which often describe nonexistent objects or fabricate facts. Existing detection methods typically operate after text generation, making intervention both costly and untimely. We investigate whether hallucination risk can instead be predicted before any token is generated by probing a model's internal representations in a single forward pass. Across a diverse set of vision-language tasks and eight modern VLMs, including Llama-3.2-Vision, Gemma-3, Phi-4-VL, and Qwen2.5-VL, we examine three families of internal representations: (i) visual-only features without multimodal fusion, (ii) vision-token representations within the text decoder, and (iii) query-token representations that integrate visual and textual information before generation. Probes trained on these representations achieve strong hallucination-detection performance without decoding, reaching up to 0.93 AUROC on Gemma-3-12B, Phi-4-VL 5.6B, and Molmo 7B. Late query-token states are the most predictive for most models, while visual or mid-layer features dominate in a few architectures (e.g., ~0.79 AUROC for Qwen2.5-VL-7B using visual-only features). These results demonstrate that (1) hallucination risk is detectable pre-generation, (2) the most informative layer and modality vary across architectures, and (3) lightweight probes have the potential to enable early abstention, selective routing, and adaptive decoding to improve both safety and efficiency.

Sai Akhil Kogilathota, Sripadha Vallabha E G, Luzhe Sun +1

Stony Brook University · Toyota Technological Institute at Chicago

Hallucination Detection in VLMs
AskPDF
Jul 8, 2026

HIVE: Understanding Post-Hallucination Reasoning in Vision Language Models

Hallucinations in vision language models (VLMs) are commonly treated as semantic errors, yet they often arise from partial or ambiguous visual evidence. Prior work mainly focuses on detecting or suppressing hallucinations at generation time, leaving the subsequent reasoning stage largely unexplored. In this work, we study Post Hallucination Reasoning (PHR), the stage in which hallucinated semantics enter the model's inference context and influence downstream predictions. To systematically investigate PHR, we introduce HIVE, Hallucination Inference and Verification Engine, an evaluation infrastructure that enables controlled comparisons between faithful and hallucinated captions. Across nine tasks and nine models, we observe structured modality dependent patterns: hallucinated captions often improve accuracy on vision language tasks, while text only tasks exhibit limited or unstable effects. Further analyses show that hallucinated cues broaden semantic coverage and reshape reasoning dynamics while preserving stable inference. These findings highlight that hallucinated semantics may influence downstream reasoning once they enter the model's inference context. Understanding this post hallucination stage is important for improving the reliability and interpretability of multimodal reasoning systems. Code is publicly available at https://github.com/hefengcs/HIVE.

Feng He, Zhenting Wang, Qifan Wang +4

Purdue University · Rutgers University · Meta AI +2

VLM EvaluationVLM HallucinationVLM Reasoning
AskPDF