大型语言模型在情绪安全分类中的比较性能跨大小和任务
Edoardo Pinzuti1,2, Oliver Tüscher1,2,3,4, André Ferreira Castro5
1Leibniz Institute for Resilience Research, Mainz, Germany.
Frontiers in artificial intelligence
|December 15, 2025
概括
大型语言模型 (LLM) 在心理健康中对处理情感内容进行评估. 较小的LLM,当微调时,显示与较大的模型可比的性能,提供保护隐私的解决方案.
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:
- 在大型语言模型 (LLM) 中评估情感内容处理对于安全的AI部署至关重要,特别是在心理健康领域.
- 目前的LLM需要对其辨别微妙情绪安全和风险类别的能力进行评估.
研究的目的:
- 为了比较不同大小的LLM在情绪安全分类任务上的表现.
- 开发和评估用于心理健康安全分类的新型数据集.
- 调查微调较小的LLM对敏感应用的有效性.
主要方法:
- 通过合并人类撰写的心理健康数据集并与ChatGPT生成的提示进行增强,构建了一个新的数据集 (>15K样本).
- 在三元 (安全/不安全/边界) 和多标签 (六类分类) 安全分类上评估了四个LLaMA模型 (1B,3B,8B,70B).
- 在零射击和少数射击学习设置中评估性能.
主要成果:
- 较大的LLM显示出更高的平均性能,特别是在复杂的多标签分类和零射击场景中.
- 轻量微调使1B参数模型能够在某些类别中与较大模型的性能相匹配,具有最小的VRAM要求 (<2GB).
- 精心调整的较小模型对于微妙的情感上下文解释和保持安全是有效的.
结论:
- 较小的,设备上的LLM是敏感的心理健康应用程序的可行,保护隐私的替代方案.
- 微调可以提高小型模型在安全关键任务中的能力.
- 这些发现对治疗性LLM开发和可扩展的AI安全对齐有重大影响.
相关概念视频
01:20Labeling Emotion
583
Emotional labeling is a cognitive process that involves identifying and naming one's emotions, such as anger, fear, happiness, or sadness. It allows individuals to recognize and express their internal emotional states, a critical aspect of emotional regulation and communication. Labeling emotions requires more than mere recognition; it also involves drawing upon memory and contextual cues to understand the current situation and apply a corresponding emotional label. For instance, feeling...
583
02:16Stereotype 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
01:19Survival Tree
369
Survival trees are a non-parametric method used in survival analysis to model the relationship between a set of covariates and the time until an event of interest occurs, often referred to as the "time-to-event" or "survival time." This method is particularly useful when dealing with censored data, where the event has not occurred for some individuals by the end of the study period, or when the exact time of the event is unknown.
Building a Survival Tree
Constructing a...
Building a Survival Tree
Constructing a...
369