PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Lexiang Xiong, Qi Li, Jingwen Ye, Xinchao Wang

arXiv:2603.15557v1cs.CV

TL;DR

VLM hallucinations are difficult to diagnose because final-output accuracy can conceal distinct, interacting failures across perception, inference, and decision-making. The paper models generation as a cognitive trajectory, projects it into an information-geometric state space, and reports state-of-the-art detection across diverse tasks with weak supervision and efficient inference. It also identifies architectural cases in which hallucinations form confident, entangled clusters that remain a challenge for diagnosis.

  • Problem

    VLM hallucinations can be multi-stage pathologies that static accuracy and monolithic detection approaches fail to distinguish, limiting trustworthy deployment.

  • Method

    The framework models generation as a normative trajectory, uses information-theoretic probes to construct a Cognitive State Space, and detects high-surprisal trajectories as geometric anomalies.

  • Results

    State-of-the-art detection is reported across POPE, MME, and MS-COCO using a single generation pass, weak supervision, and calibration resilient to up to 30% noise.

  • Takeaways & Limitations

    The framework provides stage-by-stage diagnosis of perceptual, logical-causal, and decisional failure states rather than detection alone.

  • Takeaways & Limitations

    For some architectures, hallucinations form confident clusters that intertwine with healthy states, making content errors originate from seemingly normal processes.

Abstract

from arXiv · show

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.

1 Introduction

The paper reframes VLM hallucination as a dynamic, multi-stage failure of cognitive processing rather than a static output error. It diagnoses these trajectories with information-theoretic probes embedded in an interpretable Cognitive State Space and detects hallucinations as geometric anomalies.

  • Motivation: VLM hallucinations can involve cascading perceptual and logical failures that nevertheless produce a coincidentally correct final answer.The example hallucinates a motorcycle, then contradicts that faulty evidence by answering “No.”
  • Motivation: Existing accuracy-, sampling-, and internal-state-based approaches often conflate distinct process-level failures such as perceptual drift and inferential bypass.The paper argues that hallucinations should be diagnosed within a structured model of cognition.
  • Framework: The framework models generation as a normative cognitive trajectory in which perception produces textual evidence and inference produces the final answer.Its rationality principle requires the answer to be conditionally independent of the image given the evidence, with I(A; I|Tevi) = 0.
  • Framework: Perceptual Entropy, Inferential Conflict, and Decision Entropy project uncertainty and causal inconsistency into a three-dimensional Cognitive State Space.These probes correspond respectively to perceptual instability, logical-causal failure, and decisional ambiguity.
  • Framework: Geometric abnormality is presented as an expression of high information-theoretic surprisal, turning hallucination detection into geometric anomaly detection.Nominal trajectories form stable, low-energy regions, whereas hallucinations appear as high-energy deviations from the learned manifold.
  • Results: The method achieves state-of-the-art detection across POPE, MME, and MS-COCO while using one generation pass, weak supervision, and calibration robust to up to 30% noise.The paper positions the framework as both diagnostically interpretable and practically efficient.

2 Related Work

Prior work evaluates hallucinations through final outputs, inference-time interventions, or internal-state analysis. The paper situates its framework as a process-oriented alternative to these largely static or monolithic views.

  • Output-Based Benchmarking: Hallucination benchmarks have expanded from captioned-object metrics to polling-based QA, instruction following, visual-text alignment, commonsense reasoning, and adversarial robustness.The cited suites include CHAIR, POPE, ROPE, VisIT-Bench, MMHal-Bench, HallusionBench, LM2-Bench, WYSWIR, Visual Riddles, and MAD-Bench.
  • Inference-Time Detection: Inference-time methods include contrastive decoding, hallucination-induced optimization, and automated evaluation using strong language models.These approaches target outputs driven by language priors or use contrastive and evaluator-based signals.
  • Internal-State Analysis: Internal-state research probes truthfulness representations, attention patterns, or monolithic VLM representations adapted from text-only settings.The paper characterizes this as a nascent line of inquiry distinct from its trajectory-based diagnosis.
  • Positioning: The proposed framework claims practical advantages through weak supervision and efficient inference after calibration.It requires ground-truth final answers rather than token-level annotations and uses the initial generation plus a non-autoregressive decoder pass.

3 Methodology: An Information-Geometric Framework for Diagnosing Hallucination

The framework models VLM generation as an information-geometric cognitive trajectory, diagnosing hallucinations through deviations from nominal state-space structure. It combines three probes with calibrated density modeling to score atypical cognitive states.

  • Normative Cognitive Process: A rational VLM follows I → Tevi → A, with Tevi serving as a sufficient statistic when I(A; I|Tevi) = 0.The framework treats deviations from this conditional-independence axiom as logical failures.
  • Cognitive State Space: Each generation becomes a 3D Cognitive State Vector v = [HEvi, SConf, HAns] representing perceptual, inferential, and decisional properties.These coordinates project the high-dimensional generation process onto an interpretable information manifold.
  • Diagnostic Probes: Perceptual Entropy measures uncertainty during evidence formation, Inferential Conflict measures illicit visual information leakage, and Decision Entropy measures final-answer uncertainty.Together, the probes diagnose instability at the initial stage, causal violation between stages, and residual uncertainty at the terminal stage.
  • Diagnostic Probes: Inferential Conflict estimates conditional pointwise mutual information by comparing answer-token probabilities with visual context and under visual ablation.A large positive SConf indicates direct visual information reaching the answer without mediation by generated evidence.
  • Calibration and Diagnosis: Calibration fits a Gaussian Mixture Model to standardized cognitive-state vectors from non-hallucinatory samples using only ground-truth final answers.This weak supervision avoids fine-grained token-level hallucination labels and models potentially multimodal nominal-state clusters.
  • Calibration and Diagnosis: Hallucination diagnosis assigns high scores to state vectors distant from learned high-density attractors, interpreting geometric abnormality as high surprisal.The score is the self-information of observing an atypical cognitive state, while nominal processes occupy dense regions.

4 Experiments

The experiments evaluate CAD across binary QA, comprehensive multimodal reasoning, open-ended captioning, four VLM architectures, and multiple baselines. Results report strong detection, mechanistic model-specific diagnoses, holistic gains, probe generalization, and robustness to contaminated calibration data.

  • Evaluation Protocol: CAD is evaluated on POPE, MME, and MS-COCO, alongside four VLM architectures and token-level, supervised, and other comparison baselines.POPE provides diagnostic analysis, MME tests broader multimodal tasks, and MS-COCO isolates perceptual-probe generalization in open-ended captioning.
  • Detection Performance: CAD maintains high true positive rates at FPR < 10^-2 on POPE, where baseline methods often fail.The reported low-FPR behavior is presented as evidence that modeling cognitive-process geometry improves deployment reliability over simple uncertainty thresholds.
  • Mechanistic Diagnosis: Different architectures exhibit distinct cognitive fingerprints, including Idefics2’s compact structural disorder, Llava’s separable transparent struggle, and Qwen2-VL and DeepSeek-VL2’s entangled states.The entangled pattern forms dense hallucinatory clusters that overlap healthy states, while Llava’s hallucinations show high Inferential Conflict and Decisional Ambiguity.
  • Ablation: For Idefics2, combining the probes raises detection to 0.947 AUC from individual AUCs of approximately 0.75-0.78.The ablation is presented as evidence that viewing the trajectory jointly in the 3D state space can outperform any single probe.
  • Probe Generalization: HEvi assigns significantly higher entropy to hallucinated MS-COCO captions across all models, with p ≪0.001, without task-specific tuning.The result supports standalone perceptual-probe generalization beyond structured VQA tasks.
  • Robustness: Under 0%-30% calibration contamination, Idefics2 maintains AUC above 0.91, whereas Llava degrades and Qwen2-VL and DeepSeek-VL2 remain comparatively stable.The reported differences align with each architecture’s cognitive fingerprint: compact nominal structure, high-variance conflict, or intrinsic state overlap.

5 Conclusion

The framework reframes VLM hallucinations as dynamic failures in a three-stage cognitive process and projects them into an interpretable state space for diagnosis.

  • The framework models hallucination as a dynamic failure within the model’s cognitive process rather than a static output error.
  • A three-stage cognitive trajectory is projected into an interpretable state space to support stage-by-stage diagnosis.
  • The resulting detector is state-of-the-art, single-pass, and weakly supervised.
  • The framework identifies architecture-specific cognitive fingerprints, ranging from transparent failures to deeply entangled errors.
  • The approach supports transparent auditing and the development of more reliable Vision-Language Models.

A.1 Implementation and Reproducibility Details

The experiments use a consistent Chain-of-Thought prompt and structured output format to separate evidence from final decisions, with metric definitions and uncertainty vocabulary specified for reproducibility.

  • A consistent Chain-of-Thought template elicits the three-stage cognitive process across all experiments.
  • The structured output format facilitates parsing of the evidence chain Tevi and final answer A.
  • The Chain-of-Thought template explicitly separates evidence and decision tokens.
  • Metric calculations define V as the vocabulary and I as the image.
  • The uncertainty word set VU contains linguistic uncertainty terms adapted from prior work, including space-prefixed variations.

2. Probe Calculations.

The probes quantify uncertainty in the evidence chain, visual information contribution at the decision step, and final-answer ambiguity, while a supervised linear probe provides a comparison baseline.

  • HEvi computes mean semantic binary entropy over evidence-chain tokens using their probability of uncertainty words.
  • SConf measures inferential conflict by comparing final-answer-token probabilities with and without visual features in the preceding context.
  • The visual-modality information gain is estimated through teacher-forcing replay with visual features excluded.
  • HAns is Shannon entropy over the full vocabulary distribution at the final-answer position.
  • The supervised baseline is a LogisticRegression linear probe trained on final evidence-chain hidden states with balanced data and 5-fold stratified cross-validation.

A.2 GMM Calibration and Hyperparameters

GMM calibration selects the number of nominal-cognition components with BIC and purifies calibration data by excluding inconsistent lucky guesses.

  • BIC selects the GMM component count K by sweeping K ∈[1, 10] on the calibration set.
  • The selected component counts are 5 for Llava-v1.6, 7 for Idefics2, 8 for Qwen2-VL, and 7 for DeepSeek-VL2.
  • The calibration set is filtered to remove lucky guesses where correct final answers accompany hallucinated evidence chains.
  • The coherence filter checks negation consistency and object existence when the final answer is Yes.
  • Samples failing the checks are excluded so the nominal manifold represents consistent reasoning trajectories.

B Future Work

The framework’s diagnostic capability supports future stage-specific mitigation strategies by linking detected cognitive failures to targeted interventions.

  • Future Work: Stage-specific failure localization could enable targeted mitigation strategies for distinct cognitive failures.The passage presents diagnostic localization as a basis for tailoring interventions to the failed stage.
  • Future Work: High HEvi could trigger visual re-evaluation through zooming or re-sampling.This proposed response targets failures indicated by elevated HEvi.
  • Future Work: High SConf could activate more rigorous logical verification, including self-consistency checks.This proposed response targets failures indicated by elevated SConf.
Loading 2603.15557v1…