What Do Hallucinations Reveal About Multimodal Reasoning? Diagnosing Visual Grounding Failures via Contrastive Decoding Probes

• URL: http://arxiv.org/abs/2609.16646v1
• Date: Tue, 15 Sep 2026 05:07:01 GMT
• Title: What Do Hallucinations Reveal About Multimodal Reasoning? Diagnosing Visual Grounding Failures via Contrastive Decoding Probes
• Abstract summary: We introduce SAFE, a training-free decoding framework that contrasts visually-grounded and vision-ablated generation paths.<n>This signal serves dual roles: as a practical proxy for detecting visually-ungrounded tokens, and as the basis for decoding-time penalties.<n>We argue that designing contrastive probes exemplifies a broader mission: using models as instruments for scientific understanding.
• Score: 1.9916820095972823
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: When strong multimodal models are widely available, progress requires new scientific methodologies beyond benchmark scores---using models as instruments for understanding behavior. We address this by asking: can we use large vision-language models (LVLMs) as experimental instruments for studying their own failure dynamics? Focusing on visual hallucination, we introduce SAFE, a training-free decoding framework that contrasts visually-grounded and vision-ablated generation paths to produce a token-level contrastive grounding score that identifies when the model favors linguistic priors over visual evidence. This signal serves dual roles: as a practical proxy for detecting visually-ungrounded tokens, and as the basis for decoding-time penalties. Our analysis yields three empirical observations: visual dependency decays over generation, hallucinations co-occur in temporal clusters, and early intervention reduces clustering without substantially degrading fluency. On MMHalBench, SAFE substantially outperforms all compared baselines; results elsewhere are more mixed. We argue that designing contrastive probes exemplifies a broader mission: using models as instruments for scientific understanding. Code: https://github.com/zhaozhipeng1997/SAFE_public.

Related papers

• HIVE: Understanding Post-Hallucination Reasoning in Vision Language Models [70.67112531913669]
  Hallucinations in vision language models are commonly treated as semantic errors.<n>Prior work mainly focuses on detecting or suppressing hallucinations at generation time.<n>We study Post Hallucination Reasoning (PHR), the stage in which hallucinated semantics enter the model's inference context.
  arXiv  Detail & Related papers  (2026-07-08T15:02:11Z)
• Diagnosing Visual Ignorance in Vision-Language Models [29.2851901986069]
  Vision-Language Models (VLMs) frequently rely on language priors, producing confident answers that are weakly grounded in visual evidence.<n>In this work, we study language-prior reliance from both mechanistic and behavioral perspectives.<n>We find that language-prior reliance is a systematic routing failure affecting both model internals and benchmark validity.
  arXiv  Detail & Related papers  (2026-06-05T04:16:01Z)
• SAVE: Sparse Autoencoder-Driven Visual Information Enhancement for Mitigating Object Hallucination [48.601385640941935]
  We propose SAVE, a framework that mitigates hallucination by steering the model along Sparse Autoencoder latent features.<n>A binary object-presence question-answering probe identifies the SAE features most indicative of the model's visual information processing.<n>With its simple design, SAVE outperforms state-of-the-art training-free methods on standard benchmarks.
  arXiv  Detail & Related papers  (2025-12-08T17:20:07Z)
• Grounding or Guessing? Visual Signals for Detecting Hallucinations in Sign Language Translation [13.03365340564181]
  Hallucination is a major flaw in vision-language models and is particularly critical in sign language translation.<n>We propose a token-level reliability measure that quantifies how much the decoder uses visual information.<n>Our results show that reliability predicts hallucination rates, generalizes across datasets and architectures, and decreases under visual degradations.
  arXiv  Detail & Related papers  (2025-10-21T09:13:46Z)
• Mitigating Object Hallucinations via Sentence-Level Early Intervention [10.642552315531404]
  Multimodal large language models (MLLMs) have revolutionized cross-modal understanding but continue to struggle with hallucinations.<n>We propose SENTINEL, a framework that eliminates dependency on human annotations.<n>Sentence-level Early iNtervention Through IN-domain prEference Learning can reduce hallucinations by over 90% compared to the original model.
  arXiv  Detail & Related papers  (2025-07-16T17:55:43Z)
• ViCrit: A Verifiable Reinforcement Learning Proxy Task for Visual Perception in VLMs [98.27348724529257]
  We introduce ViCrit (Visual Caption Hallucination Critic), an RL proxy task that trains VLMs to localize a subtle, synthetic visual hallucination injected into paragraphs of human-written image captions.<n>Models trained with the ViCrit Task exhibit substantial gains across a variety of vision-language models benchmarks.
  arXiv  Detail & Related papers  (2025-06-11T19:16:54Z)
• OViP: Online Vision-Language Preference Learning for VLM Hallucination [44.14029765850719]
  Large vision-language models (LVLMs) remain vulnerable to hallucination, often generating content misaligned with visual inputs.<n>We propose an Online Vision-language Preference Learning framework that dynamically constructs contrastive training data based on the model's own hallucinations.
  arXiv  Detail & Related papers  (2025-05-21T19:26:09Z)
• Generate, but Verify: Reducing Hallucination in Vision-Language Models with Retrospective Resampling [78.78822033285938]
  Vision-Language Models (VLMs) excel at visual understanding but often suffer from visual hallucinations.<n>In this work, we introduce REVERSE, a unified framework that integrates hallucination-aware training with on-the-fly self-verification.
  arXiv  Detail & Related papers  (2025-04-17T17:59:22Z)
• Treble Counterfactual VLMs: A Causal Approach to Hallucination [6.3952983618258665]
  VisionLanguage Models (VLMs) have advanced multi-modal tasks like image captioning, visual question answering, and reasoning.<n>They often generate hallucinated outputs inconsistent with the visual context or prompt.<n>Existing studies link hallucination to statistical biases, language priors, and biased feature learning but lack a structured causal understanding.
  arXiv  Detail & Related papers  (2025-03-08T11:13:05Z)
• EAGLE: Enhanced Visual Grounding Minimizes Hallucinations in Instructional Multimodal Models [54.234657224615354]
  Large language models and vision transformers have demonstrated impressive zero-shot capabilities, enabling significant transferability in downstream tasks.<n>Despite incorporating vast image and language pre-training, these multi-modal architectures often generate responses that deviate from the ground truth in the image data.<n>Current methods for mitigating hallucinations generally focus on regularizing the language component, improving the fusion module, or ensembling multiple visual encoders to improve visual representation.<n>We show that a straightforward reformulation of the original contrastive pre-training task results in an improved visual encoder that can be incorporated into the instructional multi-modal architecture without additional instructional training.
  arXiv  Detail & Related papers  (2025-01-06T00:39:31Z)
• Mitigating Hallucination for Large Vision Language Model by Inter-Modality Correlation Calibration Decoding [66.06337890279839]
  Large vision-language models (LVLMs) have shown remarkable capabilities in visual-language understanding for downstream multi-modal tasks.<n>LVLMs still suffer from generating hallucinations in complex generation tasks, leading to inconsistencies between visual inputs and generated content.<n>We propose an Inter-Modality Correlation Decoding (IMCCD) method to mitigate hallucinations in LVLMs in a training-free manner.
  arXiv  Detail & Related papers  (2025-01-03T17:56:28Z)
• AutoHallusion: Automatic Generation of Hallucination Benchmarks for Vision-Language Models [91.78328878860003]
  Large vision-language models (LVLMs) are prone to hallucinations. benchmarks often rely on hand-crafted corner cases whose failure patterns may not generalize well. We develop AutoHallusion, the first automated benchmark generation approach.
  arXiv  Detail & Related papers  (2024-06-16T11:44:43Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.