在多选题中平衡严谨性与实用性: 缓解大语言模型中的认知偏见
计算与语言
2025-04-15 v5 人工智能
摘要
本文审视了大型语言模型（LLMs）在决策过程中认知偏见的作用，挑战了消除所有偏见的常规目标。我们表明，适当地平衡认知偏见可以通过理性偏离和启发式捷径来提高决策效率。通过引入启发式调节和放弃选项（允许LLM在不确定时选择不作答），我们降低了错误率，提高了决策准确性，并优化了决策率。使用由专家合作开发的Balance Rigor and Utility（BRU）数据集，我们的发现表明，对认知偏见进行针对性检查使LLM决策更贴近人类推理，从而增强可靠性，并为未来改进提出了策略。该方法为在各种应用中利用认知偏见提高LLM的实际效用提供了新方法。
引用
@article{arxiv.2406.10999,
title = {Balancing Rigor and Utility: Mitigating Cognitive Biases in Large Language Models for Multiple-Choice Questions},
author = {Hanyang Zhong and Liman Wang and Wenting Cao and Zeyuan Sun},
journal= {arXiv preprint arXiv:2406.10999},
year = {2025}
}
备注
This work has been accepted as a full paper at the 2025 Annual Conference of the Cognitive Science Society (CogSci 2025) and will be presented in the form of a poster. The dataset and project website are available at: https://hanyangzhong.github.io/BRU-website/