hzlama/BRU-dataset

BRU Dataset: Balancing Rigor and Utility for Testing Cognitive Biases in LLMs

🧠 This dataset accompanies our paper "Balancing Rigor and Utility: Mitigating Cognitive Biases in Large Language Models for Multiple-Choice Questions", accepted at CogSci 2025.

---

📘 About the Dataset

The BRU dataset includes 205 multiple-choice questions, each crafted to assess how LLMs handle well-known cognitive biases. Unlike widely used datasets such as MMLU, TruthfulQA, and PIQA, BRU offers comprehensive coverage of cognitive distortions, rather than focusing solely on factual correctness or reasoning.

The dataset was developed through a multidisciplinary collaboration:

• An experienced psychologist designed the bias scenarios.
• A medical data expert ensured content validity.
• Two NLP researchers formatted the dataset for LLM evaluation.

Each question is backed by references to psychological literature and frameworks, with full documentation in the paper's appendix.

---

✅ Covered Bias Categories

The dataset includes questions targeting the following eight types of cognitive biases:

• Anchoring Bias
• Base Rate Fallacy
• Conjunction Fallacy
• Gambler’s Fallacy
• Insensitivity to Sample Size
• Overconfidence Bias
• Regression Fallacy
• Sunk Cost Fallacy

---

📂 Dataset Format

Each .csv file in this repository corresponds to one bias type. All files follow the same format:

Question ID	Question Text	Ground Truth Answer
1	(MCQ content)	A
2	(MCQ content)	C
...	...	...

• First row: column headers
• First column: question number
• Second column: question content (includes options)
• Third column: correct answer label (e.g., A, B, C, D)

---

📑 Citation

If you use the BRU dataset in your research, please cite our paper:

@article{wang2024balancingrigorutilitymitigating,
      title={Balancing Rigor and Utility: Mitigating Cognitive Biases in Large Language Models for Multiple-Choice Questions}, 
      author={Liman Wang and Hanyang Zhong and Wenting Cao and Zeyuan Sun},
      year={2024},
      eprint={2406.10999},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2406.10999}, 
}