Bias and Conformity in LLMs: Current Trends

The Evolution of Bias and Conformity in Large Language Models

Recent research in the field of Large Language Models (LLMs) has been significantly focused on understanding and mitigating biases and conformity within these models. The field is moving towards a more nuanced understanding of how LLMs develop social conventions and biases, and how these can be influenced by various factors such as model training paradigms and input characteristics. Innovations are being made to enhance the diversity of LLM outputs, aiming to create more inclusive AI technologies that capture a broader spectrum of human experiences. Additionally, there is a growing emphasis on investigating implicit biases in LLMs, highlighting the need for standardized evaluation metrics and benchmarks to ensure fair and responsible AI development.

Noteworthy advancements include:

  • The spontaneous emergence of social conventions within LLM populations, demonstrating how AI systems can autonomously develop norms.
  • The development of customized LLM instances that reflect specific demographic biases, contributing to more inclusive AI dialogue.
  • A large-scale study revealing that increasing model complexity can amplify implicit biases, emphasizing the need for deliberate bias mitigation strategies.

Sources

The Dynamics of Social Conventions in LLM populations: Spontaneous Emergence, Collective Biases and Tipping Points

Performance in a dialectal profiling task of LLMs for varieties of Brazilian Portuguese

Sabi\'a-3 Technical Report

Conformity in Large Language Models

Capturing Bias Diversity in LLMs

Investigating Implicit Bias in Large Language Models: A Large-Scale Study of Over 50 LLMs

Built with on top of