This research introduces a joint retrieval framework that significantly improves harmful text detection in large language models by integrating pre-trained LLMs with external knowledge sources.

• Combines the strengths of language models with structured knowledge graphs
• Demonstrates superior performance compared to single-model approaches
• Enhances robustness in detecting nuanced harmful content
• Offers practical solutions for content moderation across digital platforms

From a security perspective, this approach addresses critical AI safety concerns by improving content moderation capabilities, reducing harmful outputs, and creating more trustworthy AI systems that can better protect users across digital environments.

Improving Harmful Text Detection with Joint Retrieval and External Knowledge