Are VLMs Seeing or Just Saying? Uncovering the Illusion of Visual Re-examination

• URL: http://arxiv.org/abs/2605.15864v1
• Date: Fri, 15 May 2026 11:31:14 GMT
• Title: Are VLMs Seeing or Just Saying? Uncovering the Illusion of Visual Re-examination
• Abstract summary: Vision-Language Models (VLMs) often produce self-reflective statements like "let me check the figure again" during reasoning.<n>Do such statements trigger genuine visual re-examination, or are they merely learned textual patterns?<n>We investigate this via VisualSwap, an image-swap probing framework.
• Score: 39.713386491199884
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Vision-Language Models (VLMs) often produce self-reflective statements like "let me check the figure again" during reasoning. Do such statements trigger genuine visual re-examination, or are they merely learned textual patterns? We investigate this via VisualSwap, an image-swap probing framework: after a model reasons over an image, we replace it with a visually similar but semantically different one and test whether the model notices. We introduce VS-Bench, 800 image pairs curated from MathVista, MathVerse, MathVision, and MMMU-Pro. Experiments on Qwen3-VL, Kimi-VL, and ERNIE-VL reveal a striking failure: models overwhelmingly miss the swap, with accuracy dropping by up to 60%. Counterintuitively, thinking models are nearly 3x more vulnerable than their instructed counterparts, and scaling offers no mitigation. Multi-turn user instructions restore visual grounding, but self-generated reflective statements during continuous generation do not. Attention analysis explains why: user instructions substantially elevate attention to visual tokens, whereas self-reflection does not. Current VLMs tend to say rather than actually see when claiming to perform visual re-examination. Our code and dataset are available at the project page: https://visualswap.github.io

Related papers

• Seeing What Is Actually There: PriVE-Bench and PriVE-Tools for Counterfactual Evaluation of Agentic Visual Evidence in VLMs [14.028133377242701]
  Vision-language models (VLMs) often answer visual questions using learned language and category priors rather than grounding their predictions in the image itself.<n>We introduce PriVE-Bench, a Prior-vs-Visual Evidence Benchmark that uses paired original and counterfactual images to distinguish visually grounded answers from prior-consistent errors.<n>We also introduce PriVE-Tools, a controlled agentic-vision-inspired extension that evaluates whether tool-derived visual evidence improves grounding under the same counterfactual conflicts.
  arXiv  Detail & Related papers  (2026-07-14T16:45:02Z)
• VisualFLIP: Do Predictions Depend on Task-Critical Visual Evidence in Multimodal Reasoning? [79.44137632338062]
  When a multimodal large language model answers a visual reasoning question correctly, is the prediction actually supported by the task-critical visual evidence?<n>We introduce VisualFLIP, a paired benchmark with 1,374 images arranged as same-question pairs across cardinality, attribute, spatial, and logic tasks.<n>We evaluate 24 MLLMs with pair accuracy, which requires solving both sides of a pair, and Collapse Rate (CR), which measures how often a model that solves at least one side repeats the same non-empty answer for both images.
  arXiv  Detail & Related papers  (2026-06-05T22:06:07Z)
• How and What to Imagine? Visual Thinking in Unified Multimodal Models for Cross-View Spatial Reasoning [18.15809161363394]
  Cross-view spatial reasoning remains a weak spot for vision-language models.<n>We ask how to make visual thinking matter, and what kind of visual thinking works best.
  arXiv  Detail & Related papers  (2026-05-26T17:20:05Z)
• VisualScratchpad: Inference-time Visual Concepts Analysis in Vision Language Models [81.7389759509619]
  We introduce VisualScratchpad, an interactive interface for visual concept analysis during inference.<n>We apply sparse autoencoders to the vision encoder and link the resulting visual concepts to text tokens via text-to-image attention.<n>VisualScratchpad also provides a token-latent heatmap view that suggests a sufficient set of latents for effective concept ablation in causal analysis.
  arXiv  Detail & Related papers  (2026-03-07T20:52:09Z)
• 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)
• BLINK-Twice: You see, but do you observe? A Reasoning Benchmark on Visual Perception [67.89135437537179]
  We introduce BLINK-Twice, a vision-centric reasoning benchmark grounded in challenging perceptual tasks.<n>Instead of relying on external knowledge, our tasks require models to reason from visual content alone.<n>Compared to prior perception benchmarks, it moves beyond shallow perception and requires fine-grained observation and analytical reasoning.
  arXiv  Detail & Related papers  (2025-10-10T13:14:13Z)
• Self-Rewarding Vision-Language Model via Reasoning Decomposition [49.784411666601905]
  Vision-Language Models (VLMs) often suffer from visual hallucinations, saying things that are not actually in the image, and language shortcuts.<n>We introduce Vision-SR1, a self-rewarding method that improves visual reasoning without relying on external visual supervisions.<n>Our experiments demonstrate that Vision-SR1 improves visual reasoning, mitigates visual hallucinations, and reduces reliance on language shortcuts.
  arXiv  Detail & Related papers  (2025-08-27T08:01:03Z)
• VLMs have Tunnel Vision: Evaluating Nonlocal Visual Reasoning in Leading VLMs [18.349695067647012]
  Visual Language Models excel at complex visual tasks such as VQA and chart understanding, yet recent work suggests they struggle with simple tests.<n>We present an evaluation that tests vision-language models' capacity for nonlocal visual reasoning.<n>Our findings show that despite gains in raw visual acuity, current models lack core visual reasoning capabilities.
  arXiv  Detail & Related papers  (2025-07-04T23:15:52Z)
• v1: Learning to Point Visual Tokens for Multimodal Grounded Reasoning [27.688428439248607]
  We introduce v1, a lightweight extension that enables active visual referencing through a simple point-and-copy approach.<n>This allows the model to identify relevant image patches and copy their embeddings back into the reasoning stream.<n>Our pointing strategy lets the MLLM directly select image patches using their semantic representations as keys, keeping perceptual evidence embedded in the same space as the model's reasoning.
  arXiv  Detail & Related papers  (2025-05-24T19:30:47Z)
• Probing Visual Language Priors in VLMs [51.016683265437536]
  We introduce ViLP, a benchmark featuring deliberately out-of-distribution images.<n>Each question in ViLP is coupled with three potential answers and three corresponding images.<n>We propose a self-improving framework in which models generate new VQA data, then apply pixel-level and semantic corruptions to form "good-bad" image pairs for self-training.
  arXiv  Detail & Related papers  (2024-12-31T17:54:29Z)
• BEAF: Observing BEfore-AFter Changes to Evaluate Hallucination in Vision-language Models [20.697019266074747]
  Vision language models (VLMs) perceive the world through a combination of a visual encoder and a large language model (LLM) Recent studies show that VLMs are vulnerable to hallucination. We introduce new metrics: True Understanding (TU), IGnorance (IG), StuBbornness (SB), and InDecision (ID)
  arXiv  Detail & Related papers  (2024-07-18T12:11:12Z)
• HallusionBench: An Advanced Diagnostic Suite for Entangled Language Hallucination and Visual Illusion in Large Vision-Language Models [69.52245481329899]
  We introduce HallusionBench, a benchmark for the evaluation of image-context reasoning. The benchmark comprises 346 images paired with 1129 questions, all meticulously crafted by human experts. In our evaluation on HallusionBench, we benchmarked 15 different models, highlighting a 31.42% question-pair accuracy achieved by the state-of-the-art GPT-4V.
  arXiv  Detail & Related papers  (2023-10-23T04:49:09Z)

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.