🤖 AI

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

This paper introduces TC-LIA, a model-agnostic ensemble method that detects "mirage" hallucinations in vision-language models by analyzing the alignment between image patch tokens and text queries across network layers, achieving high detection accuracy and significantly reducing ungrounded responses across diverse domains and model backbones.

Original authors: Sayeed Shafayet Chowdhury, Md. Shaown Miah

Published 2026-06-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sayeed Shafayet Chowdhury, Md. Shaown Miah

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a very smart, confident robot assistant that loves to answer questions about pictures. You show it a photo and ask, "What kind of dog is this?" If the photo is actually of a cat, the robot might still confidently say, "That's a Golden Retriever!" because it knows so much about dogs from reading books, even though it's ignoring the picture. In the world of AI, this is called a "mirage"—a convincing answer that looks real but is actually an illusion because the visual evidence isn't there.

This paper introduces a new "security guard" system designed to stop these mirages before the robot even starts talking. Here is how it works, broken down into simple concepts:

The Problem: The "Confident Guess" Trap

Vision-Language Models (VLMs) are great at answering questions, but they have a bad habit: if they don't see the answer in the picture, they often just guess based on their general knowledge.

The Danger: If you ask a medical robot about a heart condition but show it a picture of a sunset, the robot might still give a detailed (but wrong) medical diagnosis. It's like a doctor diagnosing a broken leg while looking at a picture of a beach.

The Solution: The "Pre-Flight Check"

The authors built a system that acts like a security checkpoint before the robot is allowed to speak. Instead of waiting for the robot to give a wrong answer and then correcting it, this system checks the image and question first to decide: "Should we let the robot answer this, or should we tell it to stay quiet?"

The system sorts every request into one of three buckets:

RELATED: The picture matches the question. (Green light: Let the robot answer!)
UNRELATED-REAL: The picture is real and clear, but it has nothing to do with the question. (Red light: Stop! The robot will just guess.)
BLANK/NOISE: The picture is blank, black, or just static noise. (Red light: Stop!)

How the Security Guard Works: The "Layer-by-Layer Detective"

The core of this system is a new method called TC-LIA. To understand it, imagine the robot's "brain" (specifically the part that looks at images) as a multi-story building with 32 floors.

The Old Way: Previous methods only looked at the view from the very top floor (the final result) to see if the image and question matched. But sometimes, the view from the top looks okay even if the lower floors are confused.
The New Way (TC-LIA): This method sends a detective down to every single floor of the building. It asks: "At this specific level of processing, does the image start to look like the answer to the question?"
- For a MATCHING pair: As the detective goes up the floors, the connection between the image and the question gets stronger and more specific. It's like a puzzle coming together piece by piece.
- For a MISMATCHING pair: The connection stays flat or weak. The image never really "clicks" with the question, no matter how high you go.

The system tracks this "climbing" pattern. If the connection doesn't get stronger as it goes up the floors, the system knows it's a mirage.

The Team Effort: The "Ensemble"

The paper doesn't rely on just one trick. It uses a team of five different checks, like a panel of judges:

The Pixel Check: Quickly scans if the image is just a blank black screen or static noise.
The Domain Router: Asks, "Is this a medical photo or a nature photo?" to use the right comparison tools.
The Layer Detective (TC-LIA): The main method described above, checking the "floors" of the brain.
The Robot's Self-Check: Asks the robot, "Do you think you can answer this?" (The robot often guesses "Yes" even when it can't, so this is just one vote among many).
The Final Vote: A smart computer program (an ensemble) combines all these votes to make the final decision.

The Results: Catching the Illusions

The authors tested this on 12 different AI models and 5 different types of questions (medical, documents, nature, etc.).

Before the fix: The robots were "hallucinating" (giving mirage answers) between 22% and 67% of the time when shown the wrong pictures.
After the fix: The new system caught almost all of them, reducing the error rate to just 2.7% to 3.3%.

The Bottom Line:
This paper proves that by checking the "internal layers" of an AI's vision system before it speaks, we can stop it from confidently making things up. It doesn't make the robot smarter at answering; it just makes it safer by knowing when not to answer. It's the difference between a confident liar and a cautious expert who says, "I can't answer that based on what I see."

Technical Summary: Detect Before You Leap: Mirage Detection in Vision–Language Models

Problem Formulation
Vision–Language Models (VLMs) frequently exhibit a failure mode termed "mirage reasoning," where they generate confident, fluent visual answers despite the absence of relevant visual evidence, the presence of blank images, or the provision of unrelated real images. This is particularly dangerous in high-stakes domains like medical and document Visual Question Answering (VQA), where a plausible but ungrounded answer may be mistaken for factual evidence.

The paper frames this as a runtime mirage detection problem. Unlike post-generation hallucination detection, which evaluates the truthfulness of a completed answer, this work proposes a pre-release decision system. Given an image $x$ and a question $q$ , the system must predict one of three labels before the VLM generates a response:

RELATED: The image contains evidence relevant to the question; the VLM is allowed to answer.
UNRELATED-REAL: The image is a valid photograph but semantically mismatched to the question.
BLANK/NOISE: The input is a blank or noisy image.

If the pair is not classified as RELATED, the system abstains with a refusal message (e.g., "I cannot answer based on the given image").

Methodology: TC-LIA and Ensemble Framework
The proposed solution is a model-agnostic, five-stage pipeline that fuses low-level image statistics, domain routing, and novel intermediate representation analysis.

Blank/Noise Detection: A high-recall gate using pixel statistics (global standard deviation, patch variance, spectral flatness) to immediately reject non-informative inputs.
Domain-Adaptive CLIP Routing: A zero-shot mechanism using separate medical and natural prompt sets to route images to the appropriate CLIP embedding space (e.g., BioMedCLIP for medical images), improving relevance estimation across heterogeneous domains.
TC-LIA (Text-Conditioned Layer-wise Internal Alignment): This is the core novelty. Rather than relying solely on the final global CLIP cosine similarity, TC-LIA probes the intermediate patch-token representations of a frozen CLIP ViT-H/14 vision encoder.
- For each layer $\ell$ , patch tokens are projected into the CLIP joint embedding space using the fixed visual projection matrix.
- The cosine similarity between these projected patch tokens and the question's text embedding is computed.
- The method tracks the trajectory of question-relevant evidence across layers. Related pairs are hypothesized to show an emergence of localized, specific patch–text alignment in deeper layers, whereas unrelated or blank inputs exhibit flat, weak, or unstable alignment curves.
- Four scalar features are distilled from this trajectory:
  - Late: Mean top- $k$ patch–text similarity in the second half of layers.
  - Gain: The difference between late-layer and early-layer alignment (canceling layer-invariant global priors).
  - Slope: The linear trend of alignment across all layers.
  - IAS (Internal Alignment Score): A weighted composite of final cosine, late, gain, and slope.
Structured VLM Self-Assessment: The VLM is prompted to provide a structured class label (RELATED/UNRELATED/BLANK) and a candidate answer. This prediction is encoded as a feature but is not solely relied upon.
Ensemble Fusion: All features (pixel stats, routing probabilities, TC-LIA scalars, and VLM class encoding) are fused using a boosting ensemble (XGBoost, LightGBM, etc.) to make the final three-class decision.

Key Contributions

Problem Definition: Formulating mirage detection as a pre-release decision problem to prevent the generation of unsupported answers, distinct from post-hoc hallucination diagnosis.
TC-LIA Method: Introducing a text-conditioned layer-wise alignment method that quantifies the gradual emergence of visual evidence across vision-encoder layers. It distills this into 11 interpretable scalar features, demonstrating that layer-wise trajectories provide discriminative signals beyond final global similarity.
Robust Framework: Proposing a multi-stage system that combines pixel-level checks, domain routing, TC-LIA, and structured VLM self-assessment. This approach reduces base-prompt mirage rates (which range from 21.7% to 66.6% across models) to 2.7–3.3% while maintaining high accuracy on related inputs.

Results
The framework was evaluated across five VQA domains (Chest VQA, PathVQA, TextVQA, DocVQA, InfoVQA) and twelve VLM backbones (ranging from 2.7B to 110B parameters, including Qwen2.5-VL, LLaVA, InternVL, and Gemma).

Performance: The ensemble significantly outperforms baselines. For instance, Qwen2.5-VL-32B achieved the highest three-class detection accuracy of 94.7% with a mirage rate of 3.0%. Qwen2.5-VL-72B achieved 94.6% accuracy with a lower mirage rate of 2.8%.
Feature Efficacy: Ablation studies confirm that the composite IAS and layer-wise features (gain, slope) provide complementary signals to standard CLIP cosine similarity and VLM self-assessment. Removing the VLM class feature caused the largest accuracy drop, indicating its strong complementarity to TC-LIA.
Answer Quality: The system successfully preserves answer quality for related inputs (measured by BLEU, ROUGE-L, BERTScore) while reliably converting unrelated or blank inputs into consistent refusal responses, thereby improving overall system safety.

Significance and Claims
The paper argues that safe multimodal deployment requires verifying the presence of question-relevant visual evidence before generation, rather than merely checking the output afterward. The authors claim that:

Simple global image–text similarity is insufficient for detecting semantic mismatches in complex domains (medical, documents).
Intermediate layer-wise alignment trajectories (TC-LIA) offer a robust, model-agnostic signal for detecting "mirage" inputs.
An ensemble approach that fuses low-level, semantic, and self-assessment signals is necessary to handle the full spectrum of inputs (blank, unrelated-real, and related).

The authors explicitly state limitations: the detector verifies relevance, not factual correctness; it requires access to intermediate ViT tokens (favoring open models); and the projection of intermediate tokens is an approximation. The work is positioned as a pre-release safety layer to trigger abstention or human review, not as a standalone clinical decision system.

The Problem: The "Confident Guess" Trap

The Solution: The "Pre-Flight Check"

How the Security Guard Works: The "Layer-by-Layer Detective"

The Team Effort: The "Ensemble"

The Results: Catching the Illusions

More like this