This research introduces Representation-based Reward Modeling (RRM), a novel approach to improve safety alignment of Large Language Models while reducing computational costs.

• Addresses the distribution shift problem in reinforcement learning for LLMs without requiring expensive online sampling
• Leverages representation spaces to efficiently align models with human preferences
• Achieves comparable safety performance to more resource-intensive methods
• Provides a computationally efficient alternative for organizations with limited resources

This approach matters for security by enabling more accessible safety alignment techniques that can be applied broadly, helping prevent harmful outputs from deployed models while making safety tuning feasible for teams with constrained computational resources.

Representation-based Reward Modeling for Efficient Safety Alignment of Large Language Model