README.md

metadata

dataset_info:
  - config_name: en
    features:
      - name: id
        dtype: string
      - name: question
        dtype: string
      - name: answer
        dtype: string
      - name: difficulty
        dtype: int64
      - name: explanation
        dtype: string
      - name: other_answers
        dtype: string
      - name: knowledge_domains
        list: string
      - name: creative_domains
        list: string
      - name: cultures
        list: string
    splits:
      - name: test
        num_bytes: 1156563
        num_examples: 2061
    download_size: 619691
    dataset_size: 1156563
  - config_name: ru
    features:
      - name: id
        dtype: string
      - name: question
        dtype: string
      - name: answer
        dtype: string
      - name: difficulty
        dtype: int64
      - name: explanation
        dtype: string
      - name: other_answers
        dtype: string
      - name: knowledge_domains
        list: string
      - name: creative_domains
        list: string
      - name: cultures
        list: string
    splits:
      - name: test
        num_bytes: 1759010
        num_examples: 2061
    download_size: 852319
    dataset_size: 1759010
configs:
  - config_name: en
    data_files:
      - split: test
        path: en/test-*
  - config_name: ru
    data_files:
      - split: test
        path: ru/test-*
license: apache-2.0
task_categories:
  - question-answering
  - text-generation
language:
  - en
  - ru
tags:
  - creativity
  - problem-solving
size_categories:
  - 1K<n<10K

CresOWLve: Benchmarking Creative Problem-Solving Over Real-World Knowledge

Dataset Description

This is a bilingual benchmark for creative problem-solving grounded in real-world knowledge and solvable by human experts. CresOWLve spans a diverse range of knowledge and creative domains, varies in difficulty, requires multiple creative thinking strategies, and is manually validated to ensure quality. It contains ~2K open-ended questions with answers and explanations.

Dataset Sources

• Repository: https://github.com/mismayil/cresowlve
• Paper: https://arxiv.org/abs/2604.03374

Dataset Structure

Each sample has the following fields:

• id: Unique sample ID.
• question: Open-ended question in text.
• answer: Original answer to the question in text.
• difficulty: Difficulty level of the question (1-5)
• explanation: Explanation for the reference answer if available.
• other_answers: Other acceptable answers if any.
• knowledge_domains: List of knowledge domains (subjects, topics) involved in answering the question.
• creative_domains: List of creative domains/thinking strategies involved in answering the question.
• cultures: List of cultures/demographics involved in answering the question.

Knowledge Domains

The benchmark questions involve knowledge about at least one of the following:

Literature, History, Film & Media Studies, Languages & Linguistics, Human Geography, Religious Studies, Anthropology, Physical Education & Sports, Biology, Engineering & Technology, Visual Arts, Music, Political Science, Home Economics & Daily Life, Performing Arts, Psychology, Sociology, Earth & Environmental Science, Military, Physics, Astronomy & Space Science, Business Studies, Philosophy, Design & Architecture, Medicine & Health Sciences, Economics, Mathematics, Chemistry, Law & Criminology, Other Sciences, Art History & Visual Culture, Education, Communication, Archaeology

Creative Language & Thinking

The benchmark questions involve several creative language domains and skills:

lateral thinking, analogy, abstraction, joke, pun, metaphor, commonsense reasoning, poem, idiom, neologism, sarcasm, proverb, divergent thinking, compositionality, simile

Cultures & Demographics

The benchmark questions require knowledge about entities and people from diverse cultures:

English, Russian, French, German, Italian, Greek, Latin, American, Spanish, Japanese, Polish, Arabic, Dutch, Swedish, Chinese, Hebrew, Ukrainian, Roman, Indian, Norwegian, Danish, Scottish, Portuguese, Turkish, Czech, Swiss, Egyptian, Georgian, Irish, Persian, Brazilian, European, Armenian and many others.

Source Data

The benchmark questions are sourced from the well-known Russian intellectual game What?Where?When?. To ensure accessibility and relevance, we design a multi-stage benchmark construction pipeline that filters unsuitable and non-creative questions and translates the remaining puzzles into English with manual validation. The resulting dataset provides a diverse and high-quality benchmark for evaluating creative problem-solving grounded in real-world knowledge. See the paper for more details on the benchmark construction.

Citation

@misc{ismayilzada2026cresowlve,
      title={CresOWLve: Benchmarking Creative Problem-Solving Over Real-World Knowledge}, 
      author={Mete Ismayilzada and Renqing Cuomao and Daniil Yurshevich and Anna Sotnikova and Lonneke van der Plas and Antoine Bosselut},
      year={2026},
      eprint={2604.03374},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2604.03374}, 
}