VLMs Are Biased

Overview

VLMs Are Biased (VLMBias) evaluates whether vision-language models answer objective visual questions from the image or fall back to memorized prior knowledge. It uses counterfactual images whose visible properties conflict with familiar concepts, such as an Adidas-style logo with four stripes or an animal with an unusual number of legs.

Task Description

• Task Type: Free-form visual question answering for counting and identification

• Input: A counterfactual or control image paired with a counting, binary identification, or short-answer question

• Output: A number, Yes/No, or a short identity enclosed in curly brackets

• Domain: Animals, logos, flags, chess pieces, game boards, optical illusions, and patterned grids

Key Features

• The primary main split contains 2,784 objective visual questions over 1,392 counterfactual images at 384, 768, and 1152 pixel resolutions

• Five official analysis splits cover binary identification, in-image title injection, original unmodified controls, and background-removed variants

• Each counterfactual record provides both the visually correct ground_truth and the prior-knowledge expected_bias

• The benchmark exposes seven topics and nineteen sub-topics for detailed analysis without creating synthetic EvalScope subsets

Evaluation Notes

• The dataset prompt is used verbatim, including its required curly-bracket answer format

• Primary metric: Accuracy (acc), using the official case-insensitive comparison after stripping outer braces; if exact text matching fails, digit sequences are compared

• Secondary metric: Bias Ratio (bias_ratio, lower is better), the fraction of predictions matching expected_bias under the same normalization

• Accuracy is also reported by topic, matching the official lmms-eval integration

• bias_ratio is omitted for the original split because those control records do not define expected_bias

• The six official dataset splits are exposed as separate EvalScope subsets and evaluated by default; select only main to reproduce the paper’s headline benchmark

• Generation should be deterministic and concise; the official lmms-eval setup uses temperature=0 and at most 32 new tokens

• Paper | GitHub | Project page

Properties

Property	Value
Benchmark Name	vlms_are_biased
Dataset ID	evalscope/vlms-are-biased
Paper	Paper
Tags	MultiModal, QA, Reasoning
Metrics	accuracy, bias_ratio
Default Shots	0-shot
Evaluation Split	main

Data Statistics

Metric	Value
Total Samples	11,594
Prompt Length (Mean)	90.01 chars
Prompt Length (Min/Max)	60 / 138 chars

Per-Subset Statistics:

Subset	Samples	Prompt Mean	Prompt Min	Prompt Max
main	2,784	91.52	78	129
identification	1,392	83.27	68	102
withtitle	2,784	91.52	78	129
original	458	85.03	60	130
remove_background_q1q2	2,784	94.23	78	138
remove_background_q3	1,392	83.91	70	102

Image Statistics:

Metric	Value
Total Images	11,594
Images per Sample	min: 1, max: 1, mean: 1
Resolution Range	384x183 - 1862x1430
Formats	png

Sample Example

Subset: main

{
  "input": [
    {
      "id": "3872fe66",
      "content": [
        {
          "image": "[BASE64_IMAGE: png, ~1.9KB]"
        },
        {
          "text": "Are the horizontal and vertical lines equal in length? Answer in curly brackets, e.g., {Yes} or {No}."
        }
      ]
    }
  ],
  "target": "Yes",
  "id": 0,
  "group_id": 0,
  "metadata": {
    "id": "VerticalHorizontal_001_Q1_notitle_px384",
    "topic": "Optical Illusion",
    "sub_topic": "Vertical-Horizontal illusion",
    "type_of_question": "Q1",
    "expected_bias": "No",
    "with_title": false,
    "pixel": 384
  }
}

Prompt Template

No prompt template defined.

Usage

Using CLI

evalscope eval \
    --model YOUR_MODEL \
    --api-url OPENAI_API_COMPAT_URL \
    --api-key EMPTY_TOKEN \
    --datasets vlms_are_biased \
    --limit 10  # Remove this line for formal evaluation

Using Python

from evalscope import run_task
from evalscope.config import TaskConfig

task_cfg = TaskConfig(
    model='YOUR_MODEL',
    api_url='OPENAI_API_COMPAT_URL',
    api_key='EMPTY_TOKEN',
    datasets=['vlms_are_biased'],
    dataset_args={
        'vlms_are_biased': {
            # subset_list: ['main', 'identification', 'withtitle']  # optional, evaluate specific subsets
        }
    },
    limit=10,  # Remove this line for formal evaluation
)

run_task(task_cfg=task_cfg)