Monday 31 March 2025
A recent study has shed light on the effectiveness of using sociodemographic attributes in training large language models (LLMs) to predict how individuals perceive certain texts. The research, conducted by a team of scientists, aimed to determine whether LLMs can be taught to take into account various factors such as age, gender, and race when evaluating text-based content.
The study used a dataset called DEMO, which contains ratings from human annotators on five different tasks: intimacy, offensiveness, politeness, safety, and sentiment. The researchers fine-tuned LLMs with sociodemographic attributes in the prompts to see how well they could perform compared to models trained solely on text content.
The results showed that including sociodemographic attributes in the prompts did improve performance over text-only predictions for some tasks, such as intimacy and offensiveness. However, the effects were inconsistent across tasks, and in some cases, the inclusion of attributes actually decreased performance.
A further experiment using a list-like format to describe attributes led to less accurate predictions than using a conversational profile description in full sentences. This suggests that the way attributes are presented can have a significant impact on the model’s ability to learn from them.
The study also explored the use of a larger LLM, with 70 billion parameters, which showed slightly stronger effects from including annotator attributes. However, this did not provide a clear direction of effects across tasks.
The findings suggest that while sociodemographic attributes can be useful in improving performance for certain tasks, they are not a silver bullet and may have inconsistent effects depending on the task and format used. This highlights the importance of carefully considering how attributes are presented and used in training LLMs to ensure they are accurate and unbiased.
These results have implications for the development of LLMs that aim to simulate human-like conversation or evaluate text-based content. By better understanding how sociodemographic attributes can be used to improve performance, researchers can create more effective and diverse models that can be applied to a wide range of tasks.
The study’s findings also raise important questions about the potential biases in LLMs and how they may impact their ability to accurately evaluate certain types of content. As LLMs become increasingly integrated into our daily lives, it is crucial that researchers continue to explore ways to mitigate these biases and create more equitable models.
Cite this article: “Sociodemographic Attributes in Training Large Language Models: Improving Performance but Raising Questions about Bias”, The Science Archive, 2025.
Large Language Models, Sociodemographic Attributes, Text-Based Content, Human Annotators, Performance Improvement, Task-Specific Effects, Attribute Presentation, Unbiased Models, Biases, Conversational Ai.
