💬 NLP

MedObvious: Exposing the Medical Moravec's Paradox in VLMs via Clinical Triage

The paper introduces MedObvious, a 1,880-task benchmark demonstrating that current Vision Language Models fail to perform essential pre-diagnostic sanity checks on medical images, often hallucinating plausible narratives even when inputs are invalid or inconsistent, thereby highlighting a critical safety gap in clinical deployment.

Original authors: Ufaq Khan, Umair Nawaz, L D M S S Teja, Numaan Saeed, Muhammad Bilal, Yutong Xie, Mohammad Yaqub, Muhammad Haris Khan

Published 2026-03-25

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ufaq Khan, Umair Nawaz, L D M S S Teja, Numaan Saeed, Muhammad Bilal, Yutong Xie, Mohammad Yaqub, Muhammad Haris Khan

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 hire a brilliant, fast-talking medical student to help you read X-rays and MRI scans. This student has read every textbook in the library and can describe your symptoms in beautiful, confident, and fluent sentences. They sound like a world-class doctor.

But here is the catch: Before they can diagnose your illness, they need to make sure they are even looking at the right picture.

This is the core problem the paper "MedObvious" tackles. It argues that while AI models are getting great at talking about medicine, they are surprisingly bad at the most basic, "obvious" safety checks that a human doctor does automatically.

Here is the breakdown of the paper using simple analogies:

1. The "Medical Moravec's Paradox"

The authors coin a new term based on an old idea: Moravec's Paradox.

The Old Idea: It's hard for computers to play chess (high-level thinking), but easy for them to recognize a face (low-level perception).
The New Twist (Medical): In medicine, it's the opposite. The AI is great at the "high-level" stuff (writing a complex diagnosis report), but it fails at the "low-level" stuff that humans do without thinking, like:
- "Wait, is this an X-ray of a knee or a brain?"
- "Is this image upside down?"
- "Did the camera accidentally snap a picture of a cat instead of a patient?"

The Analogy: Imagine a tour guide who can recite the entire history of the Pyramids in perfect French, but they are leading you to the wrong pyramid entirely. The speech is perfect; the direction is wrong.

2. The "MedObvious" Test

To prove this, the researchers built a test called MedObvious.

The Setup: They showed the AI a small grid of medical images (like a 2x2 or 3x3 photo album).
The Task: They asked the AI: "In this group of pictures, is one of them a mistake? Is it the wrong body part, the wrong type of scan, or upside down? Or are they all correct?"
The "Trap": They included "Negative Controls." These were grids where every single picture was perfect and consistent. The correct answer was "No mistakes here."

The Result: Many AI models failed miserably.

When shown a perfect grid, they confidently said, "Ah, I see a tumor in the top-left picture!" (This is called a False Alarm).
They were so eager to find a problem that they hallucinated errors where none existed.
When the grid got bigger (3x3 instead of 2x2), their performance crashed, like a student getting overwhelmed when asked to compare 9 books instead of 4.

3. Why This Matters (The "Gatekeeper" Problem)

In a real hospital, a doctor doesn't just jump to a diagnosis. They first act as a Gatekeeper.

Step 1 (The Gate): "Is this the right patient? Is this the right body part? Is the image clear?"
Step 2 (The Diagnosis): "Okay, now let's look for the disease."

The paper argues that current AI models are skipping Step 1. They are jumping straight to Step 2, writing a fluent diagnosis report even when the input is garbage.

The Metaphor: Imagine a security guard at a bank. If the guard is distracted and lets a thief in because they were too busy writing a beautiful poem about the vault, the bank gets robbed. The AI is the guard who is great at writing poems but terrible at checking IDs.

4. The "Format" Trap

The researchers also found something weird: The AI's performance depended entirely on how you asked the question.

If you gave them a multiple-choice quiz ("Is it A, B, C, or D?"), they did okay.
If you asked them to just "tell me what's wrong" in their own words, they often failed or made up things.

This suggests the AI isn't actually "seeing" the error; it's just guessing based on the format of the question, like a student who memorizes the answer key but doesn't understand the math.

The Bottom Line

The paper concludes that we cannot trust medical AI yet.

Before we let AI help doctors diagnose patients, we need to teach them the "obvious" stuff first. They need to learn to say, "Wait, I can't read this," or "This image is upside down," before they try to say, "You have pneumonia."

Until AI masters these basic sanity checks, it's like handing a scalpel to someone who doesn't know how to hold it: they might have the right words, but they aren't safe to use in the real world.

1. Problem Statement: The Medical Moravec's Paradox

The paper identifies a critical safety gap in Medical Vision-Language Models (VLMs). While modern VLMs can generate fluent, coherent diagnostic narratives, they often fail basic pre-diagnostic sanity checks.

The Paradox: Extending Moravec's observation, the authors argue that tasks trivial for humans (e.g., verifying image orientation, modality, or anatomical consistency) are disproportionately difficult for VLMs, even when the models produce plausible high-level text.
The Risk: In clinical workflows, interpretation begins with verifying input validity (correct modality, anatomy, viewpoint, and integrity). If a model fails these checks, downstream diagnostic reports become clinically uninterpretable or dangerous.
Current Benchmark Limitations: Existing medical VLM benchmarks (e.g., VQA-RAD, Med-HallMark) focus on final diagnostic accuracy or text hallucination. They largely assume the input is valid, missing a critical failure mode where models hallucinate anomalies on invalid or inconsistent inputs.

2. Methodology: The MedObvious Benchmark

The authors introduce MedObvious, a benchmark specifically designed to isolate and evaluate pre-diagnostic visual triage capabilities.

A. Task Formulation

Input: Small grids of medical images ( $2 \times 2$ or $3 \times 3$ ), representing multi-view, multi-slice, or multi-series studies.
Objective: The model must determine if the grid is coherent.
- Positive Case: Identify the "outlier" panel (the one with a mismatch in modality, anatomy, viewpoint, or integrity).
- Negative Case: Confirm that no outlier exists (all panels are consistent).
Goal: To test set-level consistency rather than diagnostic reasoning.

B. Dataset Construction

Source: Curated subsets of ROCO (radiology) and Kvasir (endoscopy) to ensure modality diversity.
Generation: Tasks are template-based.
- Outlier Tasks: $n-1$ inlier panels are sampled from a reference category, and one outlier is inserted (different modality/anatomy) or an integrity violation is applied (e.g., flipped image).
- Negative Controls: All panels are sampled from the same category to test for false alarms.
Scale: 1,880 tasks total.

C. Progressive Tiers (Difficulty Levels)

The benchmark is organized into five tiers increasing in clinical specificity:

T1 (Foundation): Basic quality control (orientation, modality mismatches).
T2 (Diversity): Robustness across different data sources.
T3 (Scaling): $3 \times 3$ grids with more distractors and modality diversity.
T4 (Semantics): Anatomy and viewpoint mismatches (e.g., chest vs. abdomen).
T5 (Triage): High-saliency failures and cross-modality coherence breaks (safety cues).

D. Evaluation Protocols

To test robustness across interfaces, each task is evaluated in five formats:

Detection MCQ: Select the outlier position from options.
Detection Open: State the outlier position in free text.
Referring MCQ: Given a position, choose the correct description of the anomaly.
Referring Open: Given a position, describe the anomaly.
Visual Referring: Binary Yes/No on whether a highlighted region is the outlier.

3. Key Contributions

Formalization of the Medical Moravec's Paradox: Explicitly defining the gap between fluent diagnostic language and reliable pre-diagnostic visual verification.
MedObvious Benchmark: The first benchmark dedicated to pre-diagnostic triage, featuring 1,880 tasks, 5 progressive tiers, systematic negative controls, and 5 evaluation formats.
Comprehensive Benchmarking: Evaluation of 17 VLMs (7 general open-source, 4 medical open-source, 6 proprietary) in a zero-shot setting.

4. Key Results

The evaluation of 17 models reveals significant unreliability in pre-diagnostic gatekeeping:

Overall Performance: The best-performing model (Qwen2.5-VL-7B) achieved a mean accuracy of 63.2%, far from the reliability required for clinical safety. Human experts scored 88.4%.
False Alarms (Negative Controls): A critical failure mode is the inability to abstain. Many models hallucinate anomalies on normal, consistent inputs.
- Example: Several models showed high accuracy on "Positive" (anomaly present) cases but collapsed on "Negative" (no anomaly) cases, indicating an "always-find-something" bias.
Scaling Degradation (RQ2): Performance drops significantly when moving from $2 \times 2$ (T1/T2) to $3 \times 3$ grids (T3). This suggests models rely on salient cues rather than performing exhaustive set-level comparisons.
Format Sensitivity (RQ5): Results vary drastically based on the output format.
- Models often perform well on Multiple Choice Questions (MCQ) but poorly on Open-Ended generation (and vice versa).
- Example: Qwen2.5-VL-7B scored 75.3% on Referring MCQ but only 29.7% on Referring Open.
Proprietary vs. Open Source: Larger proprietary models (e.g., GPT-4o, Gemini) do not automatically solve the problem; they still exhibit significant false alarm rates and scaling issues.
Clinical Semantics: Models struggle with T4 (Semantics) and T5 (Triage), often confusing anatomical mismatches with pathology or failing to reject physically implausible inputs.

5. Significance and Implications

Safety Criticality: The paper argues that pre-diagnostic verification is a distinct, safety-critical capability that must be solved before VLMs can be deployed as clinical assistants or agentic tools.
Agentic Workflows: In multi-panel viewer environments (e.g., 3D Slicer, ITK-SNAP), an agent that fails to detect a flipped series or wrong modality will act on invalid data, potentially causing severe clinical errors.
Future Directions: The authors call for:
1. Calibrated Abstention: Training models to confidently say "no anomaly" when inputs are normal.
2. Systematic Comparison: Improving the ability to compare multiple images exhaustively rather than relying on single-image heuristics.
3. Interface Robustness: Ensuring model behavior is consistent across different interaction modes (MCQ vs. Open Text).

In conclusion, MedObvious demonstrates that current medical VLMs are not yet safe for autonomous clinical triage. Their fluency masks a fundamental inability to perform the basic "sanity checks" that human clinicians perform instinctively before diagnosing a patient.