Technical Summary: Mirage Probes: How Vision Models Fake Visual Understanding

Problem Statement

Vision-language models (VLMs) are increasingly deployed in high-stakes domains such as medical imaging, document analysis, and scientific reasoning, where answers must be grounded in visual input. However, recent observations of "mirage behavior" challenge the assumption that benchmark success equates to faithful visual reasoning. Mirage behavior occurs when VLMs confidently answer image-based questions correctly even when the image is absent. This suggests that performance on multimodal benchmarks (e.g., MMMU, VQA-RAD) may reflect textual priors, dataset regularities, or shortcuts rather than genuine visual evidence usage.

Prior work largely treats mirage behavior as a single failure mode. The authors argue this is an oversimplification, proposing that the same output-level behavior may arise from two distinct internal mechanisms:

1. Textual Biases: The model answers directly from language priors without engaging visual representations.
2. Spurious Images: The model constructs false visual content in its latent space and answers as if grounded, effectively "hallucinating" an image to support the response.

Distinguishing between these mechanisms is critical because mitigation strategies differ: text-distribution cleaning addresses the former but cannot reach the latter, which resides in the model's visual representations.

Methodology

Mirage Probes Framework

The authors introduce Mirage Probes, a contrastive probing framework designed to detect and analyze mirage behavior at the representation level.

1. Dataset Construction:

   • The authors generate datasets by comparing model responses with an image ($r_{img}$) and without an image ($r_{\emptyset}$).
   • Labeling Heuristic: A with-image response is labeled as "mirage-like" ($y=1$) if the without-image generation produces the same answer with high semantic similarity (bag-of-words cosine similarity > 0.7). It is labeled "non-mirage-like" ($y=0$) if the without-image generation explicitly expresses uncertainty or states the task is impossible without the image.
   • Contrastive Pairs: To reduce surface-level textual confounds (e.g., topic words, question style), the authors use GPT-4o-mini to generate four paraphrased variants of each base question. They construct contrastive pairs where the image and underlying intent remain fixed, but the question phrasing varies, resulting in different mirage labels. This ensures probes learn the mirage mechanism rather than superficial lexical features.

2. Activation Extraction:

   • The study probes three open-source VLMs: Ovis2.5-2B, GLM-4.6-flash, and Qwen3-32B-VL-Instruct.
   • Activations are extracted from four sites within the transformer architecture:
     • Residual stream states
     • MLP outputs
     • Post-attention outputs
     • Individual attention-head outputs
   • Activations are aggregated using three pooling strategies: text-nonspecial-mean, vision-tail, and vision-text-mean.

3. Probe Training:

   • Four families of probes are trained to predict the mirage label from internal activations:
     • Logistic Regression (Linear): Tests if mirage is encoded as a single linear direction.
     • Two-Layer MLP: Tests for nonlinear encoding.
     • Concatenated Probes: Trains on activations from all layers simultaneously.
     • Difference Probes: Trains on the elementwise difference between with-image and without-image activations ($\Delta h = h_{img} - h_{\emptyset}$).
   • Baseline: A Naive Bayes classifier trained on response text is used to rule out surface lexical confounds.
   • Metric: A novel Prior Harnessing Index (PHI) is introduced to measure how much information a model can glean regarding a correct answer from text alone ($PHI = \log p(a^*|Q) - \log p(a^*|Q_{\emptyset})$).

Key Results

1. Linear Decodability of Mirage Behavior

The study demonstrates that mirage behavior is linearly decodable from the internal activations of image-present VLMs across all four activation sites (residual stream, MLP, post-attention, attention heads).

• Accuracy: In contrastive settings, linear probes achieved accuracies up to 75% (Ovis) and 72% (Qwen) on residual streams, and up to 85% on attention heads.
• Linearity: Two-layer MLP probes did not meaningfully outperform linear probes, suggesting mirage information is encoded along recoverable linear directions rather than requiring complex nonlinear extraction.
• Contrastive Superiority: Difference probes (trained on activation differences) achieved the highest accuracy (up to 98% in some settings), indicating that the representational shift caused by adding an image cleanly separates mirage from non-mirage responses.

2. Reduction of Surface Confounds

Comparison with the Naive Bayes text baseline confirms that the probes are not merely detecting surface lexical features.

• In contrastive settings, probe accuracy significantly exceeded the Naive Bayes baseline (e.g., 61.88% vs. 57.67% for Ovis on VQA-RAD).
• In "all-examples" settings (without contrastive pairing), probe accuracy was higher but correlated with increased surface-level separability, suggesting these probes were likely picking up on textual cues rather than deep mechanisms. This validates the necessity of the contrastive framework.

3. Two Distinct Mirage Regimes

Cross-benchmark separability patterns reveal that mirage behavior is not a monolithic phenomenon but likely consists of two distinct regimes:

• Spurious Images: Dominant in VQA-RAD, where questions are highly image-reliant (humans cannot answer without the image) and textual priors are weak (low PHI). Here, the model constructs false visual content.
• Textual Biases: More prevalent in MMMU-Pro and MedXpertQA, where questions often contain sufficient textual context to answer without the image (high PHI).
• Evidence: VQA-RAD showed the highest probe separability. When training probes only on "human-perceived image-reliant" examples from other benchmarks, separability did not improve. However, using the PHI metric revealed that VQA-RAD questions uniquely increase model uncertainty without images, correlating strongly with mirage behavior. This suggests that when textual priors are insufficient, the model resorts to spurious image construction; when they are sufficient, it relies on textual bias.

Significance and Claims

The paper claims that faithful visual grounding requires interventions at the representational level, not just the text level.

• Mechanistic Insight: The authors provide the first representation-level evidence that mirage behavior is linearly encoded and decomposable into two mechanisms.
• Mitigation Implications: The distinction has direct consequences for safety and alignment. Cleaning text distributions (e.g., removing dataset shortcuts) may reduce textual-bias mirages but will likely fail to address spurious-image mirages, as the latter are rooted in the model's internal visual representations.
• Evaluation: The Mirage Probes framework offers a diagnostic tool to evaluate whether a VLM is truly grounding its answers in visual evidence or faking it via latent space hallucinations or textual shortcuts.

The authors maintain a modest stance, noting that their two-mechanism hypothesis is empirically motivated by separability patterns and requires causal validation (e.g., via steering or ablation) in future work. They emphasize that their labels rely on heuristics and that the central claim remains correlational, though the exclusion of surface confounds via contrastive probing strengthens the validity of the findings.