VenusBench-CAPTCHA: A Real-World CAPTCHA Screenshot–Action Benchmark for GUI Agents

Evaluation Code: https://github.com/inclusionAI/UI-Venus/tree/VenusBench-CAPTCHA

Introduction

CAPTCHA solving is a practical challenge for multimodal GUI agents because it requires more than isolated visual recognition. An agent must understand the challenge instruction, identify the relevant interface region, recognize or reason about the visual target, ground the result spatially, and produce an executable action sequence.

VenusBench-CAPTCHA is a static, offline screenshot–action benchmark designed to evaluate these capabilities in a unified format. The current release contains 219 real-world samples across 8 CAPTCHA types, covering text recognition, ordered text and icon selection, visual reasoning, one-click verification, slider manipulation, and image rotation.

Each sample provides a screenshot, a system → user conversation, the original image size, a reference action sequence, target rectangles, and an ordering constraint. Reference answers are stored separately from the model-input messages to avoid label leakage.

VenusBench-CAPTCHA contributes:

1. A real-world-first collection spanning diverse CAPTCHA appearances, resolutions, and interaction patterns.
2. A unified annotation format for single clicks, ordered multi-clicks, text entry, and dragging.
3. Pixel-space target annotations and ordering metadata for deterministic grounding and action evaluation.
4. A companion evaluation toolkit supporting local vLLM inference, OpenAI-compatible APIs, offline scoring, and visual reports.

Dataset Overview

captcha_type	Samples	Description
OCR	30	Recognize CAPTCHA text or digits and enter the answer
drag_end	30	Drag a slider or control to the required endpoint
icon_click	30	Select one or more target icons according to the instruction
image_rotation	30	Drag a rotation control until the image reaches the required orientation
oneclick_login	9	Click a one-click verification, login, or confirmation area
slider_puzzle	30	Drag a puzzle piece until it aligns with the missing region
text_click	30	Select one or more text targets, sometimes in a required order
visual_reasoning	30	Select an answer using visual semantics, correspondence, or other visual rules
Total	219

Data Structure

The repository is organized as follows:

VenusBench-CAPTCHA/
├── metadata.jsonl                 # Hugging Face Dataset Viewer metadata
├── instruction/
│   └── VenusBench-CAPTCHA.json     # Dataset annotations
├── images/
│   ├── OCR/
│   ├── drag_end/
│   ├── icon_click/
│   ├── image_rotation/
│   ├── oneclick_login/
│   ├── slider_puzzle/
│   ├── text_click/
│   └── visual_reasoning/
├── assets/
│   └── venusbench_captcha_overview.png
└── README.md

Image paths are stored relative to the annotation file. For example, ../images/slider_puzzle/15_Slide.png is resolved from instruction/VenusBench-CAPTCHA.json.

metadata.jsonl mirrors the same 219 samples in Hugging Face ImageFolder metadata format so that the Dataset Viewer can render the screenshots together with their annotations. The canonical annotation file used by the evaluation toolkit remains instruction/VenusBench-CAPTCHA.json.

Data Format

instruction/VenusBench-CAPTCHA.json is a JSON array. The following abbreviated record illustrates the schema:

{
  "images": [
    "../images/slider_puzzle/15_Slide.png"
  ],
  "messages": [
    {
      "role": "system",
      "content": "<GUI-agent and CAPTCHA task prompt>"
    },
    {
      "role": "user",
      "content": "<image>"
    }
  ],
  "image_size": [450, 850],
  "captcha_type": "slider_puzzle",
  "action_raw": "Drag(start=(86, 351), end=(245, 351))",
  "action_raw_rect": [
    [65.4072, 331.0794, 107.3765, 372.4964]
  ],
  "inorder": false
}

Field	Type	Description
images	string[]	Input image paths resolved relative to the annotation JSON
messages	object[]	Model-input conversation; all current samples contain one system message and one user message
image_size	[number, number]	Original image dimensions in [width, height] order
captcha_type	string	CAPTCHA category, matching the corresponding directory under images/
action_raw	string	Reference action sequence without the outer <action> tag
action_raw_rect	number[][]	Interaction rectangles in [x1, y1, x2, y2] pixel coordinates
inorder	boolean	Whether multiple actions or targets must be executed in annotation order

The current release uses the following reference-action forms:

Click(box=(x, y))
Click(box=(x, y)), Click(box=(x, y))
Click(box=(x, y)), Type(content='text')
Drag(start=(x1, y1), end=(x2, y2))

action_raw and action_raw_rect use original-image pixel coordinates. The supplied system prompt asks models to return coordinates in the [0, 999] space; the official evaluator converts predictions back to the original image dimensions before scoring.

Usage

No dataset-specific Python package is required. After downloading the repository, load the annotations and resolve image paths as follows:

import json
from collections import Counter
from pathlib import Path

annotation_path = Path("instruction/VenusBench-CAPTCHA.json")
samples = json.loads(annotation_path.read_text(encoding="utf-8"))

sample = samples[0]
image_path = (annotation_path.parent / sample["images"][0]).resolve()

print("samples:", len(samples))
print("categories:", Counter(item["captcha_type"] for item in samples))
print("image:", image_path)
print("reference action:", sample["action_raw"])

For model inference, validation, scoring rules, and HTML visualization, use the official evaluation implementation in the UI-Venus VenusBench-CAPTCHA branch.

The primary metric is completion-level pass@1: a single model response must contain the complete required action sequence. For full-benchmark comparison, report both micro pass@1 over all 219 samples and the unweighted macro pass@1 over all 8 CAPTCHA types.

Responsible Use and Limitations

VenusBench-CAPTCHA is intended only for authorized academic research, defensive security evaluation, model robustness analysis, and CAPTCHA usability research. It must not be used to bypass access controls on production services or for unauthorized automation, bulk account creation, spam, scraping, fraud, credential attacks, or other abuse.

This release is a static benchmark with 219 samples. It does not include live challenge backends, refresh behavior, execution feedback, or verification tokens, and it should not be used alone to claim reliable performance against production CAPTCHA systems. The categories are not perfectly balanced, screenshots vary in source and presentation, and annotations may contain noise or ambiguity.

Screenshots may contain third-party interfaces, text, brands, trademarks, or copyrighted visual elements. All third-party rights remain with their respective owners. This repository currently contains no standalone data-license file; public availability does not grant unrestricted redistribution or commercial-use rights. Users are responsible for obtaining any required permissions and complying with applicable laws, platform terms, and research-ethics requirements.

Citation

If you find this dataset useful, please cite the following paper:

@misc{venusteam2026uivenus2technicalreport,
      title={UI-Venus-2 Technical Report},
      author={Venus Team and Zhuohan Cai and Haoxing Chen and Jiaxuan Chen and Weizhi Chen and Changlong Gao and Zhangxuan Gu and Yuan Guo and Yusong Hu and Jianrong Jiang and Jianguo Li and Runze Li and Jinzhen Lin and Zhenyu Ma and Changhua Meng and Han Peng and Xinyu Qiu and Shuheng Shen and Zhongyi Shui and Weiqiang Wang and Ming Wen and Zhuoer Xu and Hang Yan and Kaiwen Yang and Ruilin Yao and Nanjun Yu and Zhengwen Zeng and Lianrui Zhang and Yunzhu Zhang and Zhe Zhao and Beitong Zhou},
      year={2026},
      eprint={2609.00028},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2609.00028},
}