FastRM is an efficient framework that automatically identifies and mitigates ungrounded responses in multimodal generative models, enhancing trustworthiness.

• Creates reference-free relevancy maps that reveal connections between inputs and outputs
• Achieves 10× faster performance than traditional gradient-based methods
• Provides confidence scores to detect potential hallucinations with 90% accuracy
• Enables real-time validation for safer deployment of vision-language models

This research addresses critical security concerns by preventing AI systems from generating misinformation and establishing greater transparency in model decision-making, ultimately building more trustworthy AI applications.

FastRM: An efficient and automatic explainability framework for multimodal generative models