🤖 AI

Revealing the Gap in Human and VLM Scene Perception through Counterfactual Semantic Saliency

This paper introduces Counterfactual Semantic Saliency (CSS), a black-box framework that reveals a significant gap between human and VLM scene perception, showing that models over-rely on object size, centrality, and saliency while under-weighting people compared to human psychophysics baselines.

Original authors: Ziqi Wen, Parsa Madinei, Miguel P. Eckstein

Published 2026-05-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ziqi Wen, Parsa Madinei, Miguel P. Eckstein

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 and a very advanced robot are looking at a complex photo of a busy street. You both describe what's happening. But here's the catch: even though you both get the "gist" of the scene, you are paying attention to completely different things.

This paper is like a detective story that tries to figure out why humans and AI see the world differently, using a new tool called Counterfactual Semantic Saliency (CSS).

Here is the breakdown of their investigation in simple terms:

1. The Problem: The "Black Box" Mystery

For a long time, scientists tried to understand how AI thinks by looking inside its "brain" (the code). But modern AI models are like giant, closed black boxes; we can't peek inside to see what they are thinking. Plus, simple tests that just ask "Is this a cat?" don't tell us how the AI decided it was a cat.

2. The New Tool: The "What If?" Game

The researchers invented a game called Counterfactual Semantic Saliency. Think of it like playing with a set of "What If?" cards.

The Setup: You show the AI a photo of a dock with a white bird and some fishing gear.
The Magic Eraser: Using high-tech AI, they digitally "erase" the bird from the photo, filling in the background so perfectly that it looks like the bird was never there.
The Test: They ask the AI to describe the new photo (without the bird).
The Measurement: They compare the description of the original photo with the description of the "erased" photo.
- If the AI's description changes drastically (e.g., from "A bird is fishing" to "Someone is fishing"), it means the AI thought the bird was super important.
- If the description barely changes (e.g., it still says "A bird is fishing" even though the bird is gone), the AI is hallucinating or ignoring the fact that the bird is missing.

By doing this for every object in the photo, they create a "heat map" of what the AI thinks matters most.

3. The Big Discovery: The "Size" Obsession

They ran this test on 19 different AI models and compared the results to 227 real humans. The results were shocking:

Humans are like detectives. We look for the story. If there's a person in the photo, we focus on them because people are usually the main characters. We ignore big, empty walls or huge rocks if they aren't doing anything interesting.
AI is like a giant, greedy magnet. It is obsessed with Size and Location.
- The Size Bias: If an object is huge, the AI thinks it's the most important thing, even if it's just a big rock in the background.
- The Center Bias: If an object is in the middle of the picture, the AI thinks it's the star of the show.
- The "Person" Blindness: Humans naturally focus on people. The AI, surprisingly, often ignores people if they are small or not in the center, focusing instead on the big, bright, or central objects.

The Analogy: Imagine a photo of a tiny ant on a giant, colorful beach ball.

You would say, "Look, there's an ant!" because the ant is the interesting part of the story.
The AI would say, "This is a giant beach ball," because the ball takes up 90% of the pixels. The AI is so focused on the "biggest thing" that it misses the actual point of the image.

4. Why This Matters (According to the Paper)

The paper concludes that the main reason AI and humans don't agree on what is important in a picture is this Size Bias. The AI has learned that "Big = Important" because it was trained on millions of photos where the main subject was usually big and in the center.

The researchers also found that the fancy "attention maps" (the usual way we try to see what AI is looking at) are often misleading. They are like a blurry, confusing scribble that doesn't match what the AI is actually saying. The "What If?" game (CSS) is the only way to get a clear answer.

Summary

The paper reveals that while AI is getting smarter at describing pictures, it is still looking at the world through a distorted lens. It sees the biggest and centrally located things as the most important, while humans see the most meaningful things (like people). Until AI learns to value "meaning" over "size," it will keep missing the forest for the trees—or in this case, missing the tiny ant on the giant beach ball.

Technical Summary: Revealing the Gap in Human and VLM Scene Perception through Counterfactual Semantic Saliency

Problem Statement

Evaluating whether large Vision-Language Models (VLMs) align with human perception in high-level semantic scene comprehension remains a significant challenge. While deep neural networks (DNNs) often achieve comparable accuracy to humans on benchmark tasks, they frequently rely on different perceptual strategies, such as over-reliance on local textures, subtle pixel patterns, or spatial relationships inconsistent with holistic human perception. Traditional "white-box" interpretability methods (e.g., gradient-based or attention maps) are often inapplicable to closed-source architectures and fail to isolate causal features. Furthermore, passive metrics cannot distinguish between correlation and causation in scene understanding. There is a critical need for a black-box, model-agnostic framework to systematically quantify object importance and measure the causal impact of visual elements on scene descriptions.

Methodology: Counterfactual Semantic Saliency (CSS)

The authors propose Counterfactual Semantic Saliency (CSS), a black-box, model-agnostic framework designed to quantify the importance of objects in scene perception by measuring the semantic shift induced by their causal ablation.

1. Counterfactual Stimuli Generation

The CSS pipeline involves three stages to generate high-fidelity counterfactual images:

Object List Generation: A large VLM (GPT-4o) generates a list of object names for a given scene, explicitly excluding background elements (e.g., sky, sea) to avoid disproportionate semantic shifts or hallucinations.
Segmentation & Mask Preprocessing: The SOTA text-prompted segmentation model, SAM3, generates segmentation masks. A four-step preprocessing pipeline ensures mask quality: removing duplicates (IoU > 95%), merging spatially proximate masks with identical labels (distance < 30 pixels) to treat structural clusters as unified entities, discarding masks occupying >30% of the image area to prevent context collapse, and applying dilation for complete coverage.
Image Inpainting: The state-of-the-art generative model Nano Banana 2 is used to remove the target objects, generating artifact-free counterfactual images. Rigorous human validation (14 annotators) ensures the target objects are successfully ablated without introducing confounding artifacts.

2. Measuring Semantic Shift

For each object $o$ in a scene $I$ , the CSS score is calculated by comparing the semantic representations of the original description and the description of the counterfactual image where $o$ is removed ( $I_{do(o=\emptyset)}$ ).

Description Generation: The target VLM generates $N=5$ independent descriptions for both the factual and counterfactual conditions.
Embedding & Similarity: Sentence embeddings are extracted using the open-source model Jasper. The CSS score is defined as one minus the mean pairwise cosine similarity between the factual and counterfactual description sets:
$CSS(o) = 1 - \frac{1}{N^2} \sum_{j=1}^{N} \sum_{k=1}^{N} \frac{E(d_j) \cdot E(d'_k)}{\|E(d_j)\| \|E(d'_k)\|}$
Interpretation: A higher CSS score indicates a larger semantic shift, implying the ablated object was more critical to the scene's comprehension. These scores are mapped to segmentation masks to create a top-down saliency map.

3. Human Baseline and Dataset

To establish a human baseline, the authors utilized the OSIE dataset (700 complex natural scenes). They filtered scenes to contain 3–6 target objects, resulting in 307 factual scenes and 1,306 counterfactual variants.

Psychophysics Study: A large-scale study was conducted with 227 independent participants via Prolific, collecting 16,289 valid textual descriptions.
Experimental Design: A between-subjects design ensured no participant viewed multiple variants of the same scene. Participants were instructed to provide a one-sentence description of the scene, mirroring the prompt used for VLMs.
Ground Truth: Human consensus was established by computing the mean pairwise similarity of a fixed subset of human responses to serve as the "ground truth" for evaluating model alignment.

Key Results

1. Pervasive Perceptual Gap

Rigorous statistical analysis reveals a systematic divergence between human cognition and VLM performance:

Top-1 Accuracy: While human-human consistency reached 73% in identifying the most critical object, all 19 evaluated VLMs (including open-source and closed-source models like GPT-5, Claude, and Gemini) fell significantly short, ranging from 57% to 65%.
Rank Correlation: The mean Kendall's $\tau$ (ordinal alignment of object importance) for human consensus was 0.58, whereas model scores ranged from 0.37 to 0.51.
Universal Failure: Every evaluated VLM failed to reach human-human consistency levels in predicting the most critical object or ranking overall object importance.

2. Divergence in Perceptual Biases

The study identified specific biases driving this gap:

Over-reliance on Low-Level Attributes: Models exhibit significantly stronger Size Bias (reliance on object area), Center Bias (reliance on image center), and Low-level Saliency Bias (reliance on contrast/color) compared to humans.
Under-utilization of Top-Down Priors: Unlike humans, who consistently prioritize People (Person Bias) in scene comprehension, models fail to replicate this prioritization, relying less on human figures to describe images.
Primary Driver: Statistical analysis isolates Size Bias as the dominant factor explaining the semantic divergence. The correlation between the divergence in Size Bias ( $\Delta r$ ) and the gap in Top-1 Accuracy ( $\Delta Acc$ ) is 0.6130 ( $p < 0.05$ ), significantly higher than other biases.

3. Limitations of White-Box Methods

The paper compares CSS against traditional white-box interpretability methods (Raw Attention, Grad-CAM, AGCAM) on an open-source VLM. The results show a remarkably low correlation (negative Kendall's $\tau$ ) between these passive attention maps and the causal CSS metric. White-box maps were often cluttered, sparse, or focused on semantically irrelevant regions, confirming their insufficiency for interpreting complex VLM architectures.

Significance and Claims

The paper claims to provide the first systematic quantification of the semantic alignment gap between modern VLMs and human observers in the context of natural scene perception.

Methodological Contribution: It introduces CSS as a novel, black-box framework that overcomes the limitations of white-box interpretability and passive metrics, enabling the isolation of causal features in closed-source models.
Empirical Contribution: It establishes a comprehensive benchmark dataset comprising 16,289 human responses across 307 scenes and 1,306 counterfactual variants, providing a robust baseline for future AI-human alignment research.
Theoretical Insight: The work demonstrates that despite scaling parameters and training data, modern VLMs still default to exaggerated, low-level visual features (specifically Size Bias) rather than top-down semantic priors (like the presence of people). This suggests that "superhuman" reasoning capabilities in VLMs may not necessarily bridge the gap in fundamental perceptual alignment and can sometimes exacerbate mismatches through over-inference or hallucination.
Safety Implications: The authors note that identifying these biases is crucial for AI safety, particularly in high-stakes environments (e.g., autonomous driving) where ignoring small but semantically critical objects in favor of large background elements could have catastrophic consequences.

The paper concludes that while VLMs show emergent alignment in isolated tasks, a pervasive gap remains in holistic scene comprehension, driven primarily by an unnatural reliance on spatial dominance (Size Bias).