This research evaluates group fairness in Large Language Model reward systems to ensure AI benefits all demographic groups equitably.

• Identifies biases in reward models that could disadvantage specific demographic groups
• Establishes new benchmarking approaches for measuring fairness across diverse populations
• Provides metrics to detect when LLMs treat certain groups differently despite similar inputs
• Proposes frameworks for developing more inclusive AI systems

From a security perspective, this work addresses critical risks of reinforcing societal biases and discrimination at scale through widely deployed AI systems, helping organizations build more ethically robust technologies.

Towards Large Language Models that Benefit for All: Benchmarking Group Fairness in Reward Models