Padding a question with filler words makes vision-language models far steadier under image corruption - but making that question more specific does the opposite.
Researchers testing Qwen3-VL and LLaVA-OneVision on the GQA and CLEVR benchmarks found that rewording 'Is there a cat?' as 'Please look carefully and answer: is there a cat?' cut answer drift by 70-81% on the 8B-parameter models when images were corrupted. Flip the change and the models get worse: turning a simple yes-or-no question into something granular, like asking what color a cup is instead of whether it exists, makes answers more likely to shift under the same corruption. The team traces both effects to the same mechanism - a question conditions the model's cross-modal attention, which acts like a frequency filter over the image. Verbose phrasing widens that filter; fine-grained phrasing narrows it onto fewer visual scales, and errors spike when the filter and the corruption land on the same frequencies.
That's a mechanistic explanation for something prompt engineers have long done on instinct: padding a request with 'look carefully' boilerplate. It means the habit isn't superstition, and it hands anyone building on top of VLMs - visual search, accessibility tools, robotics - a zero-cost accuracy lever that needs no retraining, just different wording.
So much for the push toward terse, efficient prompts - here, wordier is literally more accurate.