Vision-language models will confidently describe findings in a blank image - and do so more than half the time, according to new research.
Researchers documented a failure mode called "mirage": a vision-language model generates a plausible, confident answer to a visual question even when the image is blank, noisy, or entirely unrelated to what's being asked. Across twelve VLM architectures and five question-answering domains, baseline mirage rates ranged from 21.7% to 66.6%, meaning some models hallucinate on roughly two-thirds of invalid image inputs. The paper proposes a detection method called TC-LIA (Text-Conditioned Layer-wise Internal Alignment) that runs before the model generates a response, probing whether visual evidence matching the question actually exists in the image. Tested with that method, the best configurations brought mirage rates down to 2.8-3.0% across the Qwen2.5-VL-32B and Qwen2.5-VL-72B model families.
The stakes are highest where confident-but-wrong answers do real damage. The paper flags medical imaging and document analysis as the most dangerous domains - a model that reads X-ray findings from a blank scan isn't a curiosity; it's a patient-safety risk. The pre-generation framing is what separates this from typical hallucination filtering: rather than catching bad answers after they're written, TC-LIA decides whether to answer at all, routing the model toward abstention when visual grounding is absent.
The method still depends on a specific vision backbone - CLIP's ViT-H/14 encoder - to do its layer-wise probing, which complicates the model-agnostic claim somewhat. A 2.8% residual mirage rate also sounds reassuring until you multiply it across the millions of queries high-volume deployments handle daily. The harder question the paper leaves open: how many production VLM deployments are currently running without any pre-generation gating at all.