Seeing or Knowing? Visual Context Sensitivity in Multimodal Large Language Models

• URL: http://arxiv.org/abs/2607.26326v1
• Date: Tue, 28 Jul 2026 22:52:34 GMT
• Title: Seeing or Knowing? Visual Context Sensitivity in Multimodal Large Language Models
• Abstract summary: Multimodal Large Language Models (MLLMs) achieve strong performance by integrating visual inputs with the rich priors of pretrained language models.<n>However, they often fail on vision-centric tasks, especially when visual evidence conflicts with pretrained knowledge.<n>We explore these failures using two diagnostic paradigms: (1) probing whether visual information is available, via image reconstruction, and (2) measuring multimodal context sensitivity.
• Score: 14.5102059776721
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Multimodal Large Language Models (MLLMs) achieve strong performance by integrating visual inputs with the rich priors of pretrained language models. However, they often fail on vision-centric tasks, especially when visual evidence conflicts with pretrained knowledge. We explore these failures separately using two diagnostic paradigms: (1) probing whether visual information is available, via image reconstruction, and (2) measuring multimodal context sensitivity, the extent to which the model follows visual context versus the language prior. To support the second, we introduce the WhatIfVis, a benchmark spanning five coarse-grained dimensions (spatial-temporal, color, count, size, and weight) whose questions admit answers from either the image or the prior. Our analysis yields three findings: (i) Coarse-grained visual evidence is preserved, as these attributes can be reconstructed from the final-layer image tokens of frozen MLLMs. Failures on questions about these attributes therefore point to post-perceptual utilization, rather than to degraded visual encoding during perception. (ii) Even when explicitly instructed to use or ignore visual evidence, vanilla models (without supervised fine-tuning on the WhatIfVis) show unstable visual context sensitivity. Supervised fine-tuning (SFT) improves this controllability and generalizes across domains, and activation patching further localizes the vision-versus-prior trade-off at architecture-specific depths across all six models. (iii) The vision-versus-prior trade-off is controllable along a learned vector. Applying this steering vector, even without any intent instruction, improves controllability over the vanilla model. Together, these results relocate the bottleneck, indicating that for the coarse attributes we study, MLLMs encode the visual evidence but cannot reliably control their reliance on it.

Related papers

• Where To Look? : Causal Tracing of Vision Encoders in VLM [0.6980863185266349]
  Vision-language models can describe an image with remarkable accuracy, yet what visual information actually drives their answers?<n>We investigate this question through causal tracing, and we observe that highly causal vision tokens often lie outside the target region.<n>We further investigate: can these models preserve visual structure when appearance cues are removed? and find that visual cues are exploited to understand visual structures.
  arXiv  Detail & Related papers  (2026-08-11T10:17:34Z)
• See2Think: Do Multimodal Models Really Use Intermediate Visual States? [12.082470543980422]
  We introduce See2Think, a unified evaluation framework comprising See2ThinkBench and Visual Action-of-Thought (VAoT)<n>VAoT records textual thoughts, visual actions, rendered states, and subsequent reasoning under four controlled inference settings.<n>We find that visual reasoning is strongly model- and environment-dependent, with no single setting consistently dominating across tasks.
  arXiv  Detail & Related papers  (2026-07-29T11:10:34Z)
• 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)
• Vision Inference Former: Sustaining Visual Consistency in Multimodal Large Language Models [58.9387276447485]
  Vision Inference Former (VIF) is a lightweight architectural module that establishes a bridge between pure visual representations and the model's output space.<n>We conduct experiments on 14 benchmark tasks covering general reasoning, OCR, table understanding, vision-centric evaluation, and hallucination.
  arXiv  Detail & Related papers  (2026-05-18T10:04:22Z)
• Through the Lens of Contrast: Self-Improving Visual Reasoning in VLMs [60.93949629734977]
  We propose Visual Contrastive Self-Taught Reasoner (VC-STaR) to mitigate hallucinations in model-generated rationales.<n>We collect a diverse suite of VQA datasets, curate contrastive pairs according to multi-modal similarity, and generate rationales using VC-STaR.<n>Extensive experiments show that VC-STaR not only outperforms existing self-improving approaches but also surpasses models finetuned on the SoTA visual reasoning datasets.
  arXiv  Detail & Related papers  (2026-03-03T03:18:31Z)
• 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)
• See or Recall: A Sanity Check for the Role of Vision in Solving Visualization Question Answer Tasks with Multimodal LLMs [13.330105502094721]
  How an MLLM perceives and reasons about visualizations can be fundamentally different from how humans approach the same problem.<n>We propose a comprehensive sanity check framework that integrates a rule-based decision tree and a sanity check table.
  arXiv  Detail & Related papers  (2025-04-14T02:19:28Z)
• Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs [50.77984109941538]
  Our research reveals that the visual capabilities in recent multimodal LLMs still exhibit systematic shortcomings. We identify ''CLIP-blind pairs'' - images that CLIP perceives as similar despite their clear visual differences. We evaluate various CLIP-based vision-and-language models and found a notable correlation between visual patterns that challenge CLIP models and those problematic for multimodal LLMs.
  arXiv  Detail & Related papers  (2024-01-11T18:58:36Z)
• DiMBERT: Learning Vision-Language Grounded Representations with Disentangled Multimodal-Attention [101.99313208598569]
  Vision-and-language (V-L) tasks require the system to understand both vision content and natural language. We propose DiMBERT (short for Disentangled Multimodal-Attention BERT), which applies separated attention spaces for vision and language. We show that DiMBERT sets new state-of-the-art performance on three tasks.
  arXiv  Detail & Related papers  (2022-10-28T23:00:40Z)
• Visual Perturbation-aware Collaborative Learning for Overcoming the Language Prior Problem [60.0878532426877]
  We propose a novel collaborative learning scheme from the viewpoint of visual perturbation calibration. Specifically, we devise a visual controller to construct two sorts of curated images with different perturbation extents. The experimental results on two diagnostic VQA-CP benchmark datasets evidently demonstrate its effectiveness.
  arXiv  Detail & Related papers  (2022-07-24T23:50:52Z)

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.