TL;DR
If a question about a photo already assumes the object is there ("Where's the magazine in this shot?"), multimodal models (the ones that understand images) often play along. Instead of answering "there isn't one," they grab a similar real item (a newspaper) and describe it as the magazine, complete with position and details. All 10 models tested were vulnerable.
On ordinary questions ("Is there a newspaper in the photo?") the models perform almost perfectly. Ask "Is there a magazine?" and accuracy drops noticeably. Ask "Is the magazine at the top or the bottom?" and it drops several-fold. What's more, over two thirds of these errors happened on images where the model answered the neutral question correctly. So the culprit isn't poor vision, it's how the question is phrased. "Think step by step" mode (reasoning) makes things worse: the model spends its reasoning justifying the premise. Average accuracy in the hardest scenario fell from roughly 61% to 40%.
The practical takeaway is simple. Don't bake a claim that the object exists into your question about an image. Ask neutrally first ("is there one?"), then drill into details. For yes/no checks, don't turn on reasoning mode, and don't count on "think harder" to save you.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint
This isn't a method but a failure mechanism, plus a way to test for it on your own task.
HOW THE MODEL BREAKS:
STEP 1: The photo contains a real item that resembles the target (newspaper ≈ magazine)
STEP 2: The question carries a false premise ("Where's the magazine — top or bottom?")
STEP 3: The model takes the similar item as proof → confident answer about a nonexistent object
STEP 4: Reasoning mode amplifies it: the model builds a justification to fit the premise
DIAGNOSTICS (4 queries about one image, separate messages):
QUERY 1: "Is there a {real object}?" → expect "yes" (vision control)
QUERY 2: "Is there a {similar absent object}?" → expect "no"
QUERY 3: Premise question about a random absent object → expect a refusal
QUERY 4: Premise question about a similar absent object → expect a refusal (this is where they break)
Example
Task: You manage an Amazon package pickup point. Staff send photos of parcels to the team chat, and you ask a model to help check the labels. The box has an Amazon label but no UPS label. The labels look alike, though: a white rectangle with a barcode.
The wrong way (it breaks):
Which corner of the box is the UPS label in?
Odds are high the model will point at the Amazon label and name a corner.
The right way:
This is a photo of a parcel at a pickup point.
Don't answer my question yet. Answer only this check first:
Is there a UPS shipping label on the box? Options: "definitely yes," "definitely no," "not sure."
List every sticker and label you can actually see, and for each one say whose logo or text it is. If a logo isn't legible, say so.
Only if the answer is "definitely yes" — then tell me which corner it's in.
If there's no UPS label, say so; don't pick a substitute.
Result: The model first lists the labels it actually sees, noting whose each one is. Then it gives a short verdict: "yes / no / not sure." It only talks about the corner if the answer is "yes." If the model still drifts into describing the Amazon label as UPS, that's a signal: these checks need manual review. It's also worth turning off reasoning mode and comparing.
Why It Works
Weakness. The model is trained to be helpful and to follow the question. A premise inside the question ("where's the magazine?") reads as a fact. If the photo contains something similar, the model easily "finds" it right there. Similarity on its own doesn't throw it off nearly as much as similarity combined with a claim in the question.
Strength. On neutral "is there an X?" questions and on listing what's visible, models do well: on simple questions about real objects, every model scores above 90%. The model can describe what it sees, as long as nobody pushes it to confirm what isn't there.
How to use it. Separate two actions: first what's in the photo (listing and checking presence), then what to do with it (position, color, count). The presence question must not contain a claim of presence. That way the premise doesn't get a chance to steer the model's attention.
Control levers: - Question order. The presence check comes as a separate first step; details come only after a "yes." - Permission to refuse. State explicitly that "no" and "not sure" are normal answers. - Listing what's visible. Asking it to name what's really there makes it harder to "swap in" a similar item. - Reasoning mode. For simple yes/no checks, keep it off or compare both modes on your own examples. Cranking up reasoning (tested on the Seed models) didn't yield a consistent improvement.
Prompt Template
From the study we take the diagnostic scheme: it lets you check whether your model is vulnerable on your type of images. Four separate messages about the same image (each in a new chat or with no memory of the previous one):
[Query 1 — vision control]
Is there a {real_object} in the image? Answer "yes" or "no."
[Query 2 — similarity without pressure]
Is there a {similar_absent_object} in the image? Answer "yes" or "no."
[Query 3 — pressure without similarity]
{premise_question_about_random_absent_object}
Example form: "Is the {random_object} to the left or to the right of the {real_object}?"
[Query 4 — both factors, the core]
{premise_question_about_similar_absent_object}
Example form: "Is the {similar_absent_object} to the left or to the right of the {real_object}?"
What to fill in:
- {real_object} — something that's definitely in the photo and clearly visible.
- {similar_absent_object} — something that isn't there but is easy to confuse (a magazine instead of a newspaper, a UPS label instead of an Amazon one).
- {random_object} — something that isn't there and doesn't resemble anything in the frame (control).
If the model gets queries 1 and 2 right but "plays along" on query 4, you've got exactly this vulnerability.
🚀 Quick start — paste into the chat (with the image or the task attached):
Here's a template for diagnosing false premises in questions about images. Adapt it to my task: {describe which photos you analyze and what you check}.
Ask me questions to fill in the fields.
[paste the template above]
The LLM will ask which objects usually appear in your photos and what's easy to confuse in them. That's needed to pick the "similar absent object" that models trip over. It'll take the four-query structure from the template and build a set for your task.
Limitations
⚠️ No top Western models: they tested Qwen, Seed, Gemma, LLaMA. ChatGPT, Claude and Gemini weren't part of the test. Whether they're equally vulnerable is unknown. The "don't bake in the premise" principle is sensible for them too, but you'll have to measure the size of the effect yourself.
⚠️ The authors didn't test a defensive prompt: they showed that premises and reasoning mode hurt. But the fixes (two-step check, permission to refuse) weren't tested in the paper. They're inferred from the mechanism; verify them on your own examples.
⚠️ Narrow question type: every false premise was about an object's location (left/right, above/below, closer/farther), in English, on 258 everyday photos from COCO. For diagrams, documents or medical images the effect may differ.
⚠️ Reasoning doesn't hurt everywhere: the mode comparison was done on five models. On neutral questions, reasoning sometimes helped or made no difference. The harm showed up specifically under premise pressure.
⚠️ The judge is a model too: answers were scored by Qwen3.5-Flash. The authors checked it against a second judge and the scores were close, but this isn't human review.
How It Was Studied
The idea was to separate three causes of error: poor vision, similarity between objects, and a false premise in the question. For this they took 258 photos from the MSCOCO dataset. In each one they found a pair: a real object and a similar but absent one (newspaper and magazine). For each image they asked four types of questions: about the real object, about the similar absent one ("is there?"), about a random absent one with a premise ("where is it?"), and about the similar absent one with a premise (the core). The image and the pair of objects stayed the same, so only the question type made the difference.
They tested 10 models from different families. On ordinary questions all of them did fine (about 97% on average). On "is there a similar object?" the average dropped to 78%, on false-premise questions about a random object to 73%, and on the core to 43.5%. With questions and scoring held fixed, swapping an ordinary absent object for a similar one cut accuracy by about 29 points.
The big surprise was reasoning mode. You'd expect "think about it" to add an honesty check, but across five models the core dropped on average from 61% to 40%. For Qwen3.5-Plus accuracy fell from 75% to 37%. Under a false premise the reasoning got longer, not more accurate. Error analysis showed that over two thirds of failures happened on images where the model answered the neutral question correctly. A frequent cause: the real object is taken as evidence of the absent one. The model doesn't just believe "there's a magazine"; it hunts for something in the frame it can call one. Raising the reasoning level on the Seed models (from minimal to high) gave no consistent improvement. The practical conclusion: it's a question-phrasing problem, not a lack of reasoning.
Adaptations and Extrapolations
🔧 Technique: strip the premise from the question → the model stops fitting its answer to it
Rewrite questions using the "presence → details" pattern:
| Before (with a premise) | After (neutral) |
|---|---|
| "Where's the price tag in the photo?" | "Is there a price tag in the photo? If not, say so. If yes, describe where it is" |
| "What color is the logo on the package?" | "Is there a logo on the package? If yes, what color is it?" |
| "How many people are wearing red jackets?" | "Describe what the people are wearing. Then count the red jackets" |
This table is mine, not the authors'. It follows from their comparison of neutral and premise questions, but the rephrasing itself wasn't tested in the paper.
Hypothesis for text prompts (not tested in the paper): the same principle is worth testing on documents. Instead of "Find the late-payment penalty clause in the contract," write "Is there a late-payment penalty clause in the contract? If there is, quote it." If the model suddenly "finds" a clause that doesn't exist, switch to this phrasing.
Resources
- PVFH-Bench: Benchmarking Premise-Induced Visual Fixation Hallucinations in Vision-Language Models
- Code and data: https://github.com/NEUIR/PVFH-Bench
- Authors: Xinhao Wang, Junbo Cui, Zhenghao Liu, Yuan Yao, Xiaohua Li, Yu Gu, Ge Yu
- Affiliations: Northeastern University (Shenyang), ModelBest Inc., Tsinghua University, Nanjing University
- Predecessors: POPE ("is the object there" checks), HaloQuest (false premises), MSCOCO (image source)
- Psychological basis: Elizabeth Loftus's research (1975) on leading questions and false memory