🤖 machine learning

Multimodal Language Models Cannot Spot Spatial Inconsistencies

This paper introduces a novel benchmark for evaluating 3D motion consistency in multimodal large language models (MLLMs) by generating spatially inconsistent image pairs, revealing that current state-of-the-art models significantly underperform humans and lack a robust understanding of 3D physical structure.

Original authors: Om Khangaonkar, Hadi J. Rad, Hamed Pirsiavash

Published 2026-04-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Om Khangaonkar, Hadi J. Rad, Hamed Pirsiavash

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 are looking at a photo of a living room. Then, someone hands you a second photo of the exact same room, taken from a slightly different angle.

If you are a human, your brain instantly knows if something is "off." Maybe a chair in the second photo is floating in mid-air, or a lamp is leaning at a weird angle that defies gravity. You don't need to be a physicist to know that something is wrong.

This paper is about teaching computers to have that same "gut feeling" about the physical world.

The Problem: Computers are Great at Describing, Bad at "Feeling"

Current AI models (called Multimodal Large Language Models) are like incredibly talented art critics. You show them a picture, and they can tell you, "That's a red chair next to a blue table." They are amazing at describing what they see.

But the authors of this paper suspected that these AIs don't actually understand how the world works in 3D space. They thought the AI was just memorizing patterns of words and pixels, rather than building a mental model of how objects sit, move, and interact in reality.

The Experiment: The "Spot the Cheat" Game

To test this, the researchers created a game.

The Setup: They took real photos of rooms and used a clever trick to create "fake" inconsistencies. Imagine taking a photo of a room, then digitally cutting out a chair and pasting it back in, but from a different camera angle.
The Result: The chair now looks like it's floating or twisted in a way that is physically impossible. It's a "glitch in the matrix."
The Test: They showed pairs of these photos to both humans and top-tier AI models and asked: "Which object in the second photo breaks the rules of physics?"

The Results: Humans Win, AI Stumbles

The results were surprising and a bit worrying for the future of AI:

Humans: Were like detectives. They spotted the fake chair almost 85% of the time. It felt "off" to them immediately.
AI Models: Even the smartest, most expensive AI models (like GPT-5 or Gemini) struggled. They only got about 30-34% right. That's barely better than guessing randomly.

The Analogy: Think of it like a magic trick. A human sees the magician's hand move and knows, "He's hiding a card." The AI sees the hand move and says, "That's a hand. It is moving. It is red." The AI sees the parts, but misses the whole picture.

Why Does the AI Fail?

The researchers dug deeper and found some weird patterns:

The "Thinking" Trap: You might think, "If the AI thinks harder, it will get better." But the study found that when they forced the AI to use more "computing power" to think through the problem, it actually got worse or stayed the same. It's like a student who over-analyzes a simple math problem and ends up confusing themselves.
Brittle Brains: The AI's performance was all over the place. It might be great at spotting a floating chair in a kitchen but completely fail to notice a floating lamp in a bedroom. Humans are consistent; the AI is unpredictable.
Not Just "Glitch" Hunting: The researchers worried the AI might just be spotting digital "scratches" or editing errors (artifacts) left behind by the computer program. They tested this by adding fake scratches to the images. The AI didn't get tricked by the scratches; it was genuinely confused by the 3D geometry.

The Big Takeaway

This paper is a wake-up call. It shows that while AI can write poetry and describe a sunset beautifully, it still lacks a fundamental understanding of how the physical world is built.

It's like teaching a robot to drive by showing it millions of pictures of cars, but never letting it feel the steering wheel or understand that if you turn the wheel left, the car goes left. Until AI can "feel" the 3D world and spot when something is physically impossible, it will remain fragile and prone to making silly mistakes in the real world.

In short: The AI is a great photographer, but it's a terrible engineer. It needs to learn the laws of physics, not just the laws of language.

1. Problem Statement

The paper addresses a critical gap in the spatial reasoning capabilities of Multimodal Large Language Models (MLLMs). While MLLMs excel at describing visual scenes and semantic relationships, they often lack a robust internal model of 3D geometry and physical reality.

Core Question: Can MLLMs identify when two views of the same static scene are mutually inconsistent due to a violation of 3D motion constraints?
Hypothesis: Current SOTA MLLMs possess only surface-level 3D understanding. They may describe individual views accurately but fail to detect when a specific object's pose, scale, or occlusion is geometrically impossible given the camera movement between views.
Limitation of Existing Benchmarks: Prior methods for testing spatial reasoning often rely on novel view synthesis (e.g., Zero123, SEVA), which are computationally expensive, prone to hallucinating unrealistic geometry, or require manual annotation.

2. Methodology

A. Data Synthesis: The Cut-and-Paste Algorithm

The authors propose a lightweight, fully automatic, and scalable method to generate "spatially inconsistent" image pairs without manual annotation.

Source Data: Utilizes the Hypersim dataset (high-fidelity indoor scenes with instance-level masks, depth, and intrinsic images) and extends to DL3DV for real-world scenes.
Triplet Selection: For a given scene, three views ( $V_1, V_2, V_3$ ) are selected. An object $O$ visible in all three views is chosen, ensuring specific constraints on overlap and object size to avoid degenerate cases.
Inconsistency Generation:
- Step 1 (Inpainting): Object $O$ is removed from view $V_2$ using an inpainting model (LaMa) to create a clean background.
- Step 2 (Paste): The instance of $O$ from view $V_3$ (captured from a different camera pose) is pasted into the inpainted region of $V_2$ .
- Result: The resulting pair ( $V_1, V'_2$ ) contains a single object with a 3D pose that is geometrically incompatible with the camera transition from $V_1$ to $V_2$ , while all other scene elements remain consistent.
Efficiency: The pipeline generates >5 pairs per second on a single GPU.

B. Evaluation Protocol

Task Formulation: A forced-choice task where the model (or human) is presented with the reference view ( $V_1$ ) and the modified view ( $V'_2$ ).
Labeling: Objects in $V_1$ are labeled with letters (A–Z). The model must identify the letter corresponding to the inconsistent object.
Dataset: 615 manually verified pairs (from an auto-generated pool of 10,000+) used for evaluation.
Controls:
- Self-Pasting Baseline: Pasting the object from $V_2$ back into $V_2$ (introduces inpainting artifacts but no 3D inconsistency).
- No-Change Baseline: Unmodified $V_2$ (no artifacts, no inconsistency).
- Artifact Ablation: Varying the size of the inpainting bounding box and adding "noise" artifacts to ensure models aren't just detecting pixel-level glitches.

C. Models Evaluated

A comprehensive suite of models was tested, including:

Proprietary: GPT-5 (various reasoning budgets), GPT-4O, Gemini 2.5 (Pro/Flash), and others.
Open-Weight: Llama 3.2, Qwen3-VL, Gemma 3, Idefics, LLaVA, InternVL, and specialized spatial models like SpaceQwen.
Baselines: Random chance and human participants.

3. Key Results

A. Performance Gap

Human vs. Model: Humans achieved 84.8% accuracy. The best-performing model (GPT-5 with low reasoning budget) achieved only 34.2%.
Reasoning Budget: Increasing compute/reasoning budget (e.g., "High" vs. "Low" effort in GPT-5/Gemini) did not improve performance; in some cases, it degraded it.
Specialized Models: Models fine-tuned specifically for spatial tasks (e.g., SpaceQwen) still failed to outperform general models significantly, suggesting current training methods do not instill robust 3D geometric reasoning.

B. Variability and Brittleness

Scene Attributes: Model performance fluctuated wildly based on object depth, lighting, and physical plausibility.
- Depth: Models performed better on distant objects (>15% gap) compared to close ones, unlike humans who are stable across depths.
- Lighting: Models peaked in medium lighting, whereas humans were most accurate in bright scenes.
Object/Scene Categories: Accuracy varied significantly by object type (e.g., chairs vs. mirrors) and scene category (e.g., lounge vs. kitchen). A model might outperform another on "lamps" but fail on "sofas," indicating a lack of generalizable 3D understanding.
Error Analysis: While models often failed on the same difficult examples as humans (high IoU on wrong questions), they rarely guessed the same wrong answer. This suggests models lack a coherent alternative interpretation of the scene; their failures are unstable and model-specific.

C. Artifact vs. Geometry

Artifact Detection: Control experiments showed that models did not rely on inpainting artifacts.
- "Self-pasting" (artifacts only) performed no better than random chance.
- Adding multiple "noise" artifacts to the scene did not significantly lower performance, proving models are attempting to reason about 3D structure, not just pixel anomalies.
- Inpainting artifacts only became a confounding factor when the bounding box expansion exceeded 50%, far larger than the 5% used in the main benchmark.

4. Key Contributions

Novel Benchmark: Introduced a scalable, automated method to generate realistic spatial inconsistencies, creating a dataset of 615+ high-quality pairs for rigorous evaluation.
Diagnostic Task: Defined a simple forced-choice task that isolates projective geometry understanding from semantic description.
Empirical Evidence: Demonstrated that even the most advanced MLLMs (including GPT-5 and Gemini 2.5) possess a fragile and incomplete understanding of 3D structure, failing to detect physically impossible configurations that humans find trivial.
Analysis of Failure Modes: Provided a granular analysis showing that model failures are not random but are highly sensitive to specific scene attributes (depth, lighting, category) and lack the stability of human spatial reasoning.

5. Significance

The paper highlights a fundamental bottleneck in the current trajectory of MLLMs: the inability to ground language in physical reality.

Implication for AI Safety & Robotics: As MLLMs are deployed in robotics and autonomous systems to perceive and interact with the physical world, their inability to detect spatial impossibilities poses a significant risk.
Future Directions: The authors argue that future research must move beyond training models to simply describe the visual world. Instead, the field must prioritize developing architectures and training objectives that enforce a deep, grounded understanding of physical constraints, 3D geometry, and spatial consistency.
Benchmark Utility: The proposed dataset and evaluation protocol serve as a critical diagnostic tool for measuring progress in 3D reasoning, distinguishing between models that merely memorize visual statistics and those that truly understand spatial relationships.