Los Alamos National Laboratory Unveils Tool to Detect Hallucinations in Vision-Language AI

Summary

Researchers at Los Alamos National Laboratory have developed the Prelim Attention Score (PAS), a real-time, plug-and-play metric designed to detect hallucinations in vision-language models—AI systems that combine image analysis with large language models. Hallucinations occur when these models generate outputs describing objects or details that are absent from or inconsistent with the input image, a persistent problem affecting widely deployed AI systems. PAS acts as an internal monitor that works alongside existing vision-language models with minimal computational overhead.

The PAS system works by examining how transformer-based vision-language models allocate attention across three information sources: the input image, the text prompt, and the model's own previously generated words. By computing an attention-based score for each object mention, PAS identifies instances where models over-rely on their own generated text rather than grounding outputs in the actual image. The closer the PAS score is to zero, the less likely a hallucination has occurred. According to Manish Bhattarai, a Los Alamos computer scientist, "The system works with major existing vision-language models and requires minimal additional computational overhead, making it an efficient way to detect potential hallucinations. PAS achieves state-of-the-art accuracy in catching hallucinations."

The potential applications span high-stakes domains including medical imaging, scientific document analysis, engineering diagrams, and remote sensing, where unsupported visual claims could significantly impact downstream decisions. The Los Alamos team is presenting PAS at the Computer Vision and Pattern Recognition (CVPR) 2026 conference in Denver this month, which is sponsored by the IEEE and Computer Vision Foundation.

• Applications include medical imaging, scientific document analysis, engineering diagrams, and remote sensing where hallucinations could affect critical decisions
• Research being presented at CVPR 2026, offering developers a practical path toward safer and more trustworthy multimodal AI systems

Editorial Opinion

PAS represents a meaningful step toward practical AI safety in production systems. Rather than attempting to solve hallucinations at the architectural level—a still-unsolved problem—Los Alamos offers a pragmatic monitoring solution that integrates with existing models. The focus on attention patterns is technically sound and the low computational overhead makes this deployable at scale. However, detection without correction has limited value; the real test will be whether developers actually integrate PAS into workflows and whether flagged hallucinations lead to meaningful interventions.