种族,民族及其对大型语言模型偏见的影响
medRxiv : the preprint server for health sciences
|January 16, 2026
概括
这项研究揭示了大型语言模型 (LLM) 如何处理种族和种族,发现通过神经元干预的偏见缓解显示出有限的成功,这表明AI模型中更深层次的代表性问题.
科学领域:
- 人工智能的人工智能
- 自然语言处理自然语言处理.
- 计算语言学 计算语言学
背景情况:
- 大型语言模型 (LLM) 越来越多地用于医疗保健等敏感领域.
- 现有的研究强调了结果的差异,但缺乏对LLM内部机制的洞察力.
- 了解LLM如何代表人口统计特征对于减轻偏见至关重要.
研究的目的:
- 调查大型语言模型代表和运行种族和民族的内部机制.
- 在LLM架构中分析人口信息的分布和功能.
- 评估有针对性的干预措施对减少偏见的影响.
主要方法:
- 使用两个公共数据集 (毒性生成,临床叙述理解) 分析了三个开源LLM.
- 应用可重现的解释性管道,结合探测,神经元级归因和有针对性的干预.
- 检查人口统计线索如何影响模型行为和内部表示.
主要成果:
- 人口信息分布在内部模型单元中,具有显著的跨模型可变性.
- 一些单位从预训练数据中表现出与刻板印象相关的学习.
- 同样的人口线索可以触发不同的模型行为,干预措施产生不完整的偏差减少.
结论:
- 法学士课程对种族和民族的内部表示是复杂的,并且在不同模型中存在差异.
- 针对特定神经元的偏差缓解策略显示出有限的有效性,这表明更深层次的表征性挑战.
- 需要进一步的研究,以更有系统的方法来解决LLMs中的偏见.
相关概念视频
Stereotypes, Prejudice, and Discrimination
95.0K
Humans are very diverse and although we share many similarities, we also have many differences. The social groups we belong to help form our identities (Tajfel, 1974). These differences may be difficult for some people to reconcile, which may lead to prejudice toward people who are different. Prejudice is a negative attitude and feeling toward an individual based solely on one’s membership in a particular social group (Allport, 1954; Brown, 2010). Prejudice is common against people who...
95.0K
Language and Cognition
716
Language serves as a bridge between ideas and communication, influencing how individuals perceive and interact with the world. Psychologists have long debated whether language shapes thought or vice versa. This discussion gained grip with Edward Sapir and Benjamin Lee Whorf in the 1940s, who proposed that language determines thought, a concept known as linguistic determinism. They suggested that the vocabulary and structure of a language influence how its speakers think and perceive reality.
716
Bias
7.2K
Bias refers to any tendency that prevents a question from being considered unprejudiced. In research, bias occurs when one outcome or answer is selected or encouraged over others in sampling or testing. Bias can occur during any research phase, including study design, data collection, analysis, and publication.
In statistics, a sampling bias is created when a sample is collected from a population, and some members of the population are not as likely to be chosen as others (remember, each member...
In statistics, a sampling bias is created when a sample is collected from a population, and some members of the population are not as likely to be chosen as others (remember, each member...
7.2K
Surveys
16.6K
Often, psychologists develop surveys as a means of gathering data. Surveys are lists of questions to be answered by research participants, and can be delivered as paper-and-pencil questionnaires, administered electronically, or conducted verbally. Generally, the survey itself can be completed in a short time, and the ease of administering a survey makes it easy to collect data from a large number of people.
16.6K
Motivational Bias
313
Cognitive bias results from limitations in thinking and information processing, leading to systematic errors in judgment. Conversely, motivational bias stems from personal desires or emotions, causing distortions in perception to align with self-interest. Motivational bias influences how individuals perceive and attribute causes to events, often shaped by personal needs, goals, and self-esteem preservation. This bias can distort judgment, leading to inaccurate assessments of success, failure,...
313
Stereotype Content Model
15.3K
The Stereotype Content Model (SCM) was first proposed by Susan Fiske and her colleagues (Fiske, Cuddy, Glick & Xu, 2002; see also Fiske, 2012 and Fiske, 2017). The SCM specifies that when someone encounters a new group, they will stereotype them based on two metrics: warmth—or that group’s perceived intent, and how likely they are to provide help or inflict harm—and competence—or their ability to carry out that objective. Depending on the warmth-competence...
15.3K


