You show a vision-language model (VLM) a geometry problem. The image contains a right triangle with three labeled side lengths. The model says: "Let me take another look at the image to confirm..." — and gives the correct answer.
But when you secretly swap the image for a different triangle with different side lengths that looks similar, the model's answer barely changes.
It said "let me look again." It did not look.
Shi, Yang, Wu, Jin, Shui, Berg-Kirkpatrick, and Ma recently published a paper on arXiv (2605.15864), accepted as an ICML 2026 Spotlight, doing something simple but deeply unsettling: they built an image-substitution probing framework called VisualSwap.
🔀 The VisualSwap Experiment
The setup is straightforward. First, ask a VLM a question with visual dependency — what's in this image? During its answer, the model might say "let me double-check this image."
At the moment it should "re-confirm," the experimenters quietly swap the image for another one — visually similar but semantically different. For example, two shopping receipts with different amounts. Or two similar triangles with different dimensions.
Then they check whether the model's post-"re-check" answer reflects the change.
The results are striking. Across Qwen3-VL, Kimi-VL, and ERNIE-VL, models almost universally missed the swap. Accuracy dropped by up to 60%.
More counterintuitively: thinking models — those trained to perform multi-step reasoning before answering — were more easily fooled than standard instruction models, by a gap of nearly three times. And larger model scale did not make models harder to deceive.
🧠 But What Does "Looking" Actually Mean?
The key point here is not "VLMs can be fooled." The key point is that when a VLM is asked to "confirm," it produces a piece of self-reflective text — "let me look at the image again" — but this text does not trigger any new processing of visual information. It is a learned linguistic pattern, not a genuine visual re-examination behavior.
How do you prove this? The paper analyzes the model's attention allocation. When the user (explicitly, in the conversation) says "please re-check this image," the model's attention on visual tokens rises significantly — it is actually looking. But when the model itself produces a "let me look again" statement, attention shows no comparable shift. The eyes (attention) did not move.
> This may resonate intuitively with anyone who has used GPT-4o or Gemini in multi-turn conversations. Sometimes a model says "let me recalculate" mid-reasoning and then gives exactly the same wrong answer as before. You wonder: did it really recalculate? This paper demonstrates that — for vision, at least — much of the time, it genuinely did not.
📉 Why Are Thinking Models More Vulnerable?
The paper finds that thinking models — those trained to produce an internal chain-of-thought before answering — are more sensitive to this illusion. The likely reason: thinking models have longer contexts and must manage more textual material, and mentions of "re-checking the image" get buried in the reasoning. Instruction models respond better to "please check" instructions in the input because they lack large amounts of internal reasoning text to distract them.
Scale does not help either. Larger models show no better visual re-examination ability. This indicates the problem is not "the model isn't smart enough" — it's that the architecture contains no mechanism that enforces visual re-examination.
🤷 What Remains Unclear
First, how were the 800 swap-image pairs in VS-Bench selected? If the swapped images sit close to the originals in the models' visual feature spaces — i.e., they genuinely "look" similar to the models — then missing the swap might partly reflect reasonable generalization rather than pure visual neglect. The paper describes the swaps as "visually similar but semantically different," but whether that similarity threshold aligns with the evaluated models' feature spaces is unclear.
Second, the paper finds that user instructions work but model self-reflection does not. Yet there is an important difference between the two: a user instruction arrives in a new conversational turn, often accompanied by a fresh image input. Self-reflection is generated autoregressively within the same reasoning step. Does this mean it is not that "self-reflective text is ineffective," but rather that the model's retrieval ability within the same context is inherently limited? The paper does not disentangle these.
Third, when do self-reflective statements appear during reasoning? Does saying "let me look again" early in visual reasoning have a different effect than saying it late? The paper performs attention analysis but no temporal ablation. The same sentence at different moments might have completely different effects.
Still, the paper's core warning is large and clear: when a VLM says it "looked" at something, don't believe it. It may just be uttering a sentence it learned sounds natural.
---
References
1. Shi, C., et al. (2026). *Are VLMs Seeing or Just Saying? Uncovering the Illusion of Visual Re-examination*. arXiv:2605.15864 [cs.CV]. https://arxiv.org/abs/2605.15864 (ICML 2026 Spotlight) 2. OpenAI. (2024). *GPT-4V(ision) System Card*. Technical Report. 3. Bai, J., et al. (2025). *Qwen3-VL: A Versatile Vision-Language Model*. arXiv preprint. 4. Team Kimi. (2025). *Kimi-VL Technical Report*. arXiv preprint. 5. Augenstein, I., et al. (2024). *Seeing the Unseen: Visual Hallucination in Large Vision-Language Models*. ACL 2024.