[Paper Review] A Comprehensive Survey of Bias in LLMs: Current Landscape and Future Directions

Rajesh Ranjan, Shailja Gupta+1 more|arXiv (Cornell University)|Sep 24, 2024

Artificial Intelligence in Law9 citations

TL;DR

This paper surveys biases in large language models, detailing bias types, sources, impacts, mitigation strategies, and future research directions.

ABSTRACT

Large Language Models(LLMs) have revolutionized various applications in natural language processing (NLP) by providing unprecedented text generation, translation, and comprehension capabilities. However, their widespread deployment has brought to light significant concerns regarding biases embedded within these models. This paper presents a comprehensive survey of biases in LLMs, aiming to provide an extensive review of the types, sources, impacts, and mitigation strategies related to these biases. We systematically categorize biases into several dimensions. Our survey synthesizes current research findings and discusses the implications of biases in real-world applications. Additionally, we critically assess existing bias mitigation techniques and propose future research directions to enhance fairness and equity in LLMs. This survey serves as a foundational resource for researchers, practitioners, and policymakers concerned with addressing and understanding biases in LLMs.

Motivation & Objective

• Motivate the need to understand biases in LLMs for safer deployment across applications.
• Categorize biases across multiple dimensions including sources and impacts.
• Synthesize findings from current research on bias and mitigation in LLMs.
• Propose future directions to improve fairness and equity in LLMs.

Proposed method

• Systematic categorization of biases into multiple dimensions.
• Synthesis of existing research findings on bias types, sources, and impacts.
• Critical assessment of current bias mitigation techniques.
• Discussion of real-world implications and policy considerations.
• Proposal of future research directions to enhance fairness in LLMs.

Experimental results

Research questions

• RQ1 What are the principal bias types present in LLMs and their sources?
• RQ2 How do biases in LLMs impact real-world applications and users?
• RQ3 What mitigation strategies exist for LLM biases and how effective are they across contexts?
• RQ4 What future directions are recommended to improve fairness and equity in LLMs?

Key findings

• Biases in LLMs span multiple dimensions, including type, source, and impact.
• Current research provides a synthesis of bias types, their origins, and consequences in applications.
• Mitigation techniques exist but require critical assessment and broader evaluation across contexts.
• The paper outlines future directions to enhance fairness and equity in LLM deployment.