analiza aiqbench

Los Alamos National Laboratory's Prelim Attention Score (PAS) establishes a new state-of-the-art benchmark for detecting hallucinations in vision-language models by monitoring attention patterns during token generation. This real-time, plug-and-play metric offers developers a low-overhead solution to verify output fidelity against input images, which is critical for deploying trustworthy multimodal AI in high-stakes domains like medical imaging. By pinpointing the origin of fabricated details within the generation process, PAS provides a practical mechanism to enhance the reliability of autoregressive systems without significant computational cost.