Researchers at Los Alamos National Laboratory have introduced the Prelim Attention Score (PAS), a new method for detecting hallucinations in vision-language models.
These AI systems combine image recognition with language generation to answer questions or describe visual content. They are widely used in applications ranging from document analysis to image interpretation.
One of the biggest challenges facing these systems is hallucination. This happens when an AI model describes objects, details, or events that do not actually appear in the image. Such errors can reduce confidence in AI-generated results and create problems in situations where accuracy is important.
PAS helps address this issue by monitoring how a vision-language model generates each word in its response. The system determines whether the model relies on actual image data or its own previously generated text, helping flag outputs that may contain hallucinations.
The new PAS system helps identify these mistakes as they occur. Instead of checking the final answer alone, PAS monitors how the AI generates each word in a response. This allows it to determine whether the model is relying on visual information or on its own previously generated text.
According to researcher Manish Bhattarai, PAS acts as an internal monitoring tool for AI systems. He explained that the technology works with major vision-language models and imposes very little additional computational burden. The team reported that PAS achieves state-of-the-art performance in detecting hallucinations while remaining efficient enough for practical use.
READ ALSO: Cornell’s ShadowDraw AI Transforms Everyday Object Shadow Into Creative Artwork
How Los Alamos PAS Works
Many modern vision-language models use transformer architectures. These deep-learning systems rely on a mechanism known as attention, which helps determine which information is most important when generating a response. Attention can focus on image features, user prompts, or words already produced by the model.
PAS examines these attention patterns during response generation. It measures how much influence previously generated words have on the next word being produced. This helps reveal when the model starts depending more on its own text than on the original image.
The system assigns a numerical score for object mentions in an AI-generated response. A score closer to zero indicates a lower chance that the statement is a hallucination. Higher scores suggest the model may be drifting away from the visual evidence provided in the image.
Researchers say this process provides a simple but effective warning signal. Rather than requiring major changes to existing AI models, PAS uses information that the models already generate internally. This makes it easier to integrate into current workflows and AI products.
Why Reliable Vision AI Matters
Vision-language models are becoming important across many industries. They are used to analyze photographs, technical diagrams, scanned documents, satellite imagery, and other visual data. As adoption grows, ensuring that these systems provide accurate information has become a major priority.
The Los Alamos team believes PAS can support reliability checks in several high-value fields. These include medical imaging, scientific research, engineering analysis, and remote sensing applications. In such environments, unsupported visual claims can influence decisions and create costly errors.
READ ALSO: Gemini North Captures Crystal Ball Nebula’s Stunning Final Stage 1,500 Light-Years Away
Researcher Xuan Nhat Hoang said PAS helps pinpoint the exact moment when a model begins to rely too heavily on its own generated content. By reading signals already produced by the AI, the system offers a low-overhead method for improving trustworthiness. This approach helps users better understand whether a response is grounded in actual visual evidence.
The timing of the research is significant, as organizations continue to deploy multimodal AI systems at scale. Businesses and institutions increasingly require tools that can evaluate AI reliability alongside performance. PAS represents a step toward greater transparency in how AI systems reach their conclusions.
The research team is presenting the technology at the 2026 Computer Vision and Pattern Recognition Conference in Denver, one of the leading global events in artificial intelligence and computer vision. As vision-language AI becomes more deeply integrated into critical workflows, tools such as PAS are expected to play an important role in making AI outputs more dependable and easier to verify.
