Datasets:

Warrior0302
/

CAP-Bench

license: cc-by-4.0
language:
  - en
task_categories:
  - text-generation
  - question-answering
  - other
task_ids:
  - open-domain-qa
pretty_name: 'CAP-Bench: Cross-Site Browser-Agent Benchmark'
size_categories:
  - n<1K
annotations_creators:
  - expert-generated
language_creators:
  - expert-generated
source_datasets:
  - original
multilinguality:
  - monolingual
tags:
  - benchmark
  - agent-evaluation
  - agents
  - browser-agents
  - web-agents
  - cross-site
  - gui
  - llm-evaluation
  - agent-as-a-judge
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test.csv

CAP-Bench

A Scalable Benchmark for Evaluating Cross-Site Browser Agents with Complex Actions and Perception.

CAP-Bench evaluates browser agents on Cross-site workflows, complex Actions, and challenging visual Perception. The full benchmark contains 420 tasks across 108 real-world websites in 24 functional domains. Each task requires on average 7 complex execution operations and 4 perception challenges, substantially exceeding the difficulty of prior browser-agent benchmarks.

This Hugging Face dataset hosts the public split (192 tasks). A held-out private split (228 tasks) is reserved for contamination-resistant leaderboard evaluation; see the GitHub repo for submission details.

🔗 Links


🏠 Project page	https://warriorxu0302.github.io/CAP-Bench/
🏆 Leaderboard	https://warriorxu0302.github.io/CAP-Bench/leaderboard.html
📖 Paper	arXiv preprint coming soon
💻 Code	https://github.com/WarriorXu0302/CAP-Bench

Why CAP?

Real web browsing is harder than current benchmarks suggest. Modern web applications present two intertwined challenges that prior evaluations largely overlook:

Execution complexity — completing a task often requires non-trivial UI interactions (dragging map regions, manipulating date-range sliders, navigating nested dropdowns) that go well beyond simple clicking and typing.
Perception complexity — users must interpret rich visual content including charts, tables, images, and dynamically rendered elements to extract actionable information.

Compounding this, realistic workflows demand such interactions across multiple websites. CAP targets all three dimensions and pairs them with a verifiable agent-as-a-judge evaluation framework that scores agents against fine-grained execution and perception checkpoints rather than only the final answer.

Dataset Structure

Data fields

Field	Type	Description
`task_id`	`string`	Unique identifier of the form `task-<6-hex>`.
`instruction`	`string`	Natural-language user request describing a multi-step, cross-site task. Includes desired output format.

Future revisions will additionally expose the structured metadata used in the paper (covered action / perception points, involved websites, functional clusters). Track https://github.com/WarriorXu0302/CAP-Bench for release notes.

Data splits

Split	Tasks	Notes
`test` (this repo)	192	Public subset — use for development and reporting.
(held out)	228	Private subset for leaderboard evaluation. Submit results via the GitHub leaderboard issue template.

Loading

from datasets import load_dataset

ds = load_dataset("Warrior0302/CAP-Bench")
print(ds)
# DatasetDict({
#     test: Dataset({ features: ['task_id', 'instruction'], num_rows: 192 })
# })

print(ds["test"][0]["task_id"])
# 'task-035e06'
print(ds["test"][0]["instruction"][:120])
# "I'm writing a sci-fi script about AI ethics and want to systematically gather creative source material..."

Statistics

192 public tasks (this split)
108 real-world websites covered (full benchmark)
24 functional domains (full benchmark) — e-commerce, travel booking, academic search, real estate, code hosting, video streaming, news, healthcare, and more
~1,279 characters per instruction on average; min 601, max 2,495

Evaluation

CAP ships with CAP-Eval, a verifiable agent-as-a-judge framework that:

Generates a per-task Python evaluation script from each task's covered points.
Materializes a hierarchical rubric tree of independently checkable execution / perception leaves.
Has a judge agent extract structured information from the agent's answer and verify each claim against live webpages.

Reported metrics:

Partial Completion — mean root-node score across tasks.
Success Rate — fraction of tasks scoring 1.0.
Complex-A — mean leaf score over action nodes.
Complex-P — mean leaf score over perception nodes.

See src/evaluate/ in the code repo for the full evaluator and CLI (run_eval.py).

How to submit to the leaderboard

Run your agent on all 192 public tasks (use the instruction field as the user prompt).
Save each agent answer as answers/<your_agent>/<task_id>/answer_<n>.md.
Run python run_eval.py --agent_name <your_agent> from the code repo, or report self-computed metrics if you have your own evaluator.
Open a leaderboard issue with your numbers, agent description, and a link to evaluation logs.

License

The dataset is released under Creative Commons Attribution 4.0 (CC BY 4.0). The accompanying code (CAP-Eval, construction pipeline) is released under Apache 2.0.

Citation

If CAP-Bench helps your work, please cite:

@misc{cap2026,
    title  = {CAP: A Scalable Benchmark for Evaluating Cross-Site Browser Agents
              with Complex Actions and Perception},
    author = {Xu, Zejun and Chen, Taiyi and Li, Jin and Gu, Yongtong and Cheng, Qi
              and Lv, Aixuan and Zhu, Shuai and Zhu, Pengfei and Yang, Kaichen
              and Sun, Boyu and Yang, Yixian and Xie, Mulong and Liu, Xin
              and Li, Dagang and Ma, Xiaoteng and Wang, Hongru},
    year   = {2026},
    note   = {Preprint, arXiv link forthcoming}
}

Acknowledgements

The CAP-Eval framework extends Mind2Web2 (Gou et al., 2025), originally released under the MIT License. See the GitHub repo's NOTICE file for details.