Mirage Probes: How Vision Models Fake Visual Understanding

Vision-language models (VLMs) can produce confident answers without relevant visual evidence, a failure mode known as mirage reasoning (Asadi et al., 2026). To that end, we study pre-release mirage detection: deciding whether a VLM answer should be released or withheld. Our model-agnostic method, Text-Conditioned Layer-wise Internal Alignment (TC-LIA), tracks question-image alignment across the layers of a frozen CLIP ViT-H/14 encoder, summarizing patch-text alignment by final similarity, late-layer top-k alignment, early-to-late gain, and slope. TC-LIA is purely unsupervised (fixed projections, fixed scoring weights, no labels, no training) and already delivers strong detection independently. Additionally, when combined with blank/noise detection, domain routing, and VLM self-assessment, it forms an ensemble whose supervised training improves performance but is an optional add-on. On 19,004 samples spanning ten VQA domains, fourteen state-of-the-art VLMs exhibit 57.3-75.0% base mirage rates. Our proposed TC-LIA alone cuts this to 7.5% with 83.5% Related/Unrelated/Blank-Noise classification accuracy, and the ensemble reaches 84.3-88.4% accuracy with 5.9-7.2% mirage rates (best joint result: 88.4% accuracy, 6.4% mirage rate). Notably, an ensemble trained on a single backbone transfers well to unseen backbones, with the best-transferring source staying within 1.2% accuracy points of per-backbone training across thirteen held-out VLMs.

Vision-Language Models (VLMs) frequently rely on language priors, producing confident answers that are weakly grounded in visual evidence. While this behavior is widely observed, its internal mechanisms and its impact on benchmark evaluation remain insufficiently understood. In this work, we study language-prior reliance from both mechanistic and behavioral perspectives. Internally, we combine counterfactual layer replacement with supervised layer-wise MLP probing to trace how ground-truth visual semantics and language-prior semantics compete across the language decoder. Our analysis reveals a multi-stage bottleneck: intermediate layers often fail to effectively retrieve visual information, while later layers can further suppress surviving visual signals in favor of text-space biases. Externally, we introduce a progressive visual decay metric based on multi-step Gaussian blurring, which identifies instances whose answers remain invariant even as visual content is increasingly destroyed. Across twelve visual question-answering benchmarks and three representative VLMs, we find that a substantial fraction of examples remain answerable under severe or total visual obfuscation, indicating that current benchmarks can inadvertently reward visual ignorance. These findings demonstrate that language-prior reliance is a systematic routing failure affecting both model internals and benchmark validity. Finally, we outline critical pathways for future research, highlighting the necessity of designing training distributions and evaluation protocols built on structurally isolated or counterfactual data to enforce genuine cross-modal grounding.

Benchmark accuracy is often implicitly assumed to reflect grounded visual understanding in vision-language models (VLMs), yet it remains unclear to what extent such scores truly reflect reliance on visual evidence. Motivated by a surprising observation that removing a substantial fraction of image tokens only degrades model performance very slightly on a widely used hallucination benchmark, we systematically investigate this mismatch in a set of open-source VLMs. Our analysis spans multiple levels of granularity, spanning global visual degradation, localized occlusion, question reformulation, answer-space expansion, and decision-level analyses beyond standard accuracy. We further complement these behavioral results with a layer-wise analysis of vision-token geometry. Throughout the experiments, we find that although VLMs do incorporate visual input, their predictions are less sensitive to the loss of fine-grained visual evidence that standard accuracy should have suggested. Even when the final prediction remains unchanged, the model's internal support for the correct answer may already be weakened. We further complement a representation-level analysis, which shows increasing similarity among visual tokens in deeper layers, providing a possible explanation for our findings. Together, these results suggest that current benchmarks are not sufficient to reliably evaluate fine-grained visual grounding in VLMs.