基于用户画像和对话历史的LLM行为适应评估数据集
收藏资源简介:
本研究提出了一种评估大型语言模型(LLM)如何根据用户的年龄、职业和教育水平等社会人口统计学特征调整其反应的框架。为了实现这一目标,研究团队构建了一个合成数据集,该数据集将对话历史与独特的用户画像进行配对,并使用来自价值观调查模块(VSM 2013)的问题来探索价值表达。数据集的创建过程涉及使用多代理管道,将来自精心策划的模拟数据集中的用户画像与对话历史配对,从而生成包含社会人口统计学属性的对话数据。该数据集旨在解决LLM在处理不同格式的用户属性时是否能够保持行为一致性的问题,并评估LLM在反映用户期望方面的能力。此外,研究还发现,大多数模型在年龄和教育水平等人口统计学特征发生变化时,会调整其表达的价值,但不同模型在不同输入格式之间的一致性存在差异。具有更强推理能力的模型在格式之间的一致性方面表现更好,这突出了推理在稳健的社会人口统计学适应中的重要性。
This study proposes a framework for evaluating how Large Language Models (LLMs) adjust their responses based on users' sociodemographic characteristics such as age, occupation, and educational attainment. To achieve this goal, the research team constructed a synthetic dataset that pairs conversation histories with unique user profiles, and uses questions from the Values Survey Module (VSM 2013) to explore value expression. The dataset creation process employs a multi-agent pipeline to pair user profiles from a carefully curated simulation dataset with conversation histories, thereby generating conversational data containing sociodemographic attributes. This dataset aims to address the question of whether LLMs can maintain behavioral consistency when handling user attributes in different formats, and evaluate the ability of LLMs to reflect user expectations. Furthermore, the study found that most models adjust their expressed values when sociodemographic characteristics such as age and educational level change, but there are differences in consistency across different input formats among different models. Models with stronger reasoning capabilities perform better in terms of cross-format consistency, which highlights the importance of reasoning in robust sociodemographic adaptation.

- 1Evaluating LLM Adaptation to Sociodemographic Factors: User Profile vs. Dialogue History南洋理工大学, 新加坡 · 2025年



