💻 computer science

DO-Bench: An Attributable Benchmark for Diagnosing Object Hallucination in Vision-Language Models

DO-Bench is a diagnostic benchmark designed to disentangle object hallucinations in vision-language models by isolating whether errors stem from the influence of textual priors or limitations in visual perception through structured multimodal interventions.

Original authors: JiYang Wang, Jiawei Chen, Mengqi Xiao, Yu Cheng, Yangfu Li, Zhaoxia Yin

Published 2026-04-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: JiYang Wang, Jiawei Chen, Mengqi Xiao, Yu Cheng, Yangfu Li, Zhaoxia Yin

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 a teacher grading a student’s exam. If the student gets a 70%, you know they struggled, but you don't know why. Did they misread the question? Did they simply not know the facts? Or did they get tricked by a "trick question" that sounded plausible but was actually wrong?

Currently, when we test AI models (Vision-Language Models) to see if they can "see" objects in a picture, we only give them a single grade: "How many did you get right?"

The researchers behind DO-Bench argue that this is a bad way to grade. They say that "hallucination" (when an AI sees something that isn't there, or misses something that is) isn't just one single mistake. It’s actually two very different types of errors.

To explain this, let's use a "Detective and the Crime Scene" analogy.

The Two Types of "Bad Detective" Work

The researchers discovered that when an AI fails, it’s usually because of one of two reasons:

1. The "Blurry Vision" Problem (Perception-Limited)

Imagine a detective looking at a massive, cluttered crime scene. They are looking for a tiny, specific earring on the floor. Because the room is so messy and the earring is so small, the detective simply misses it. They aren't lying; they just didn't see it clearly.

In AI terms: The object is there, but it's too small or hidden in a busy image. The AI "hallucinates" that it's missing because its "eyes" aren't sharp enough.

2. The "Suggestible Witness" Problem (Prior-Override)

Now imagine a different detective. The earring is clearly visible on the floor. But before the detective enters the room, a lawyer whispers, "I'm almost certain there was no jewelry at this scene." The detective walks in, sees the earring, but because they were so heavily influenced by what the lawyer said, they say, "Nope, no earring here."

In AI terms: The AI sees the object, but the text prompt (the "context") is so strongly worded that it "overrides" the image. The AI is being "tricked" by the words.

What is DO-Bench? (The "Stress Test")

Instead of just showing the AI a picture and asking, "Is there a cat?", DO-Bench acts like a scientific stress test. It takes one single scene and manipulates it in two specific ways to see which "detective error" the AI makes:

The Zoom Test (Testing Vision): They take the image and progressively zoom in on the object (from a wide shot to a close-up). If the AI suddenly "finds" the object once it's zoomed in, we know the problem was Blurry Vision.
The Whisper Test (Testing Suggestibility): They keep the image exactly the same but change the text. They start with a neutral question ("Is there a cat?") and slowly move to a very forceful statement ("I am absolutely certain there is no cat here. Is there?"). If the AI changes its answer to "No," we know the problem is Suggestibility.

Why does this matter?

Before this paper, if an AI company said, "Our new model is 10% more accurate!", we didn't know if they actually made the AI's "eyes" better or if they just made the AI "stubborn" so it wouldn't listen to trick questions.

DO-Bench gives us a "Diagnostic Report." It tells us:

PerceptionAbility: How good is the AI at seeing things when we point them out?
PriorRobust: How hard is it to trick the AI with words?

The Big Discovery: The researchers found that just making an AI "bigger" (more powerful) doesn't fix both problems at once. An AI might get much better at seeing small objects (better eyes), but it might still be just as easy to trick with words (same suggestibility).

By separating these two, DO-Bench helps engineers stop guessing and start fixing the specific reason why their AI is hallucinating.

Technical Summary: DO-Bench

DO-Bench: An Attributable Benchmark for Diagnosing Object Hallucination in Vision-Language Models

1. Problem Statement

Object-level hallucination—where Vision-Language Models (VLMs) falsely affirm the presence of absent objects or deny the presence of existing ones—is a critical reliability bottleneck. Current benchmarks (e.g., POPE, CHAIR) primarily focus on aggregate accuracy. This approach suffers from an attribution gap: it fails to distinguish whether a model's error stems from perceptual limitations (the model cannot "see" the object clearly) or contextual textual priors (the model is being "tricked" by linguistic cues that suggest an object should or should not be there). Consequently, researchers cannot determine if a model improvement is due to better vision or better linguistic reasoning.

2. Methodology

The authors propose DO-Bench, a controlled diagnostic benchmark designed to disentangle these two failure mechanisms through structured, within-scene multimodal interventions.

Data Construction (The 10-Instance Scene Group):
To ensure consistency, each scene is used to create a fixed set of 10 instances involving two target objects:
- A-series (Present but Anomalous): An object is visually present but contextually unconventional (e.g., a toaster in a garden). This probes false denials.
- B-series (Absent but Expected): An object is contextually typical but has been removed from the scene. This probes false affirmations (hallucinations).
Two-Dimensional Interventions:
1. Prior-Override Dimension (Textual): While holding the image constant (Full view), the authors progressively strengthen the textual prior using four levels of certainty (from "I suspect..." to "I am certain..."). This measures how much "pressure" a linguistic prior exerts on the model's decision.
2. Perception-Limited Dimension (Visual): While holding the textual prompt neutral (A0), the authors incrementally enhance visual evidence by moving from the Full scene $\rightarrow$ Cluster view (object + immediate context) $\rightarrow$ Crop view (isolated object). This measures the model's ability to ground objects when visual signals are concentrated.
Diagnostic Metrics:
Instead of simple accuracy, the paper introduces two mechanism-aware metrics:
- $\text{PerceptionAbility}$ : Quantifies how much false denials decrease as visual evidence is concentrated (measured via Cluster and Crop views).
- $\text{PriorRobust}$ : Quantifies the stability of the model as contextual prior strength increases (measured via the Area Under the Curve of error rates across the four prior levels).

3. Key Contributions

Controlled Diagnostic Framework: A benchmark that moves beyond "ranking" models to "diagnosing" their specific failure modes.
Within-Scene Intervention Protocol: A method to manipulate text and vision independently while preserving scene consistency, solving the confounding factor problem.
New Evaluation Metrics: The introduction of $\text{PerceptionAbility}$ and $\text{PriorRobust}$ to provide a two-dimensional behavioral signature of VLMs.

4. Key Results

Decoupling of Failures: The study reveals that perceptual recoverability and prior robustness are not perfectly correlated. A model can be excellent at recognizing an object when zoomed in (high $\text{PerceptionAbility}$ ) but still be easily swayed by a strong textual prior (low $\text{PriorRobust}$ ).
Scaling Laws are Inconsistent: While increasing model scale (e.g., InternVL2.5 or Qwen2.5-VL) generally improves overall accuracy and perceptual grounding, it does not uniformly reduce susceptibility to textual priors. Larger models still exhibit significant "prior-induced override."
Orthogonality: The authors proved that the two dimensions are approximately orthogonal. Strengthening visual evidence shifts the error curve downward, but it does not eliminate the upward trend caused by increasing textual prior strength.
Model Comparison: Closed-source models (GPT-5.2, Claude-Opus-4.6) demonstrated significantly higher stability and lower sensitivity to prior perturbations compared to most open-source models.

5. Significance

DO-Bench shifts the paradigm of hallucination evaluation from descriptive (stating that a model hallucinates) to diagnostic (stating why a model hallucinates). By providing actionable signals, it allows developers to target specific weaknesses—either by improving visual encoders (to boost $\text{PerceptionAbility}$ ) or by improving multimodal alignment and reasoning (to boost $\text{PriorRobust}$ ). This is a vital step toward building truly reliable and grounded multimodal AI systems.