Detect Before You Leap: Mirage Detection in Vision-Language Models

Vision-language models (VLMs) can produce confident visual answers even when the required visual evidence is missing, blank, or unrelated to the question. This failure mode, recently described as mirage (Asadi et al., 2026), is especially concerning in medical and document VQA, where visually ungrounded answers may be mistaken for image-based evidence. We study pre-release mirage detection: given an image-question pair, determine whether a VLM's answer should be released or the system should abstain before the answer reaches the user. We propose Text-Conditioned Layer-wise Internal Alignment (TC-LIA), a model-agnostic method that probes patch-token representations across the layers of a CLIP ViT-H/14 vision encoder. The key idea is to project layer-wise image patch tokens into the final CLIP embedding space and measure their similarity with the question embedding, tracking whether question-relevant visual evidence emerges across vision layers. TC-LIA summarizes this trajectory using final image-text cosine similarity, late-layer top-k patch-text alignment, early-to-late gain, and layer-wise slope. These features are combined with pixel-statistic-based blank/noise detection, zero-shot domain routing, and structured VLM self-assessment in an ensemble classifier. We evaluate on 19,004 samples spanning ten VQA domains, including medical imaging, document understanding, scene text, reasoning, and video understanding, with four input conditions: RELATED, IN-DOMAIN UNRELATED, OUT-OF-DOMAIN UNRELATED, and BLANK/NOISE. Across fourteen VLMs, base prompt mirage rates range from 57.3-75.0%. TC-LIA alone reduces the mirage rate to 7.5% at 83.5% accuracy; ensemble fusion further improves performance to 84.5-88.4% accuracy with 5.7-7.2% mirage rates. The best result, obtained with Qwen2.5-VL-72B, achieves 88.4% accuracy and a 6.4% mirage rate.