Researchers at Los Alamos National Laboratory have developed a new tool called the Prelim Attention Score, or PAS, to help detect when a vision-language model’s output is grounded in the image and when it may be relying too heavily on its own generated text.
The system is being viewed as a means to enhance vision-language model safety and trustworthiness.
“The PAS is a real-time, plug-and-play metric that acts as an internal monitor for the AI,” said 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, offering developers a practical path toward safer and more trustworthy multimodal AI systems,” he added.
Most commonly used vision-language models are autoregressive. They generate each new token, or word, partly by relying on the words they have already produced.
While this process helps the model form coherent responses, it can also cause the system to lean too strongly on its own previous output rather than the image itself.
PAS monitors a vision-language model’s prediction of each token. By doing so, it helps identify where the model is drawing information from and where hallucinations are likely to occur. The tool then presents a score that alerts users to the possible presence of hallucinations in the output.
Many autoregressive vision-language models are based on transformer architectures, a class of deep-learning neural networks that use attention patterns to weigh information as they generate an output.
The Los Alamos team studied how these models attend to the image, the text prompt and the model’s own preliminary generated words.
When PAS is integrated into a vision-language model workflow, it can run alongside the model as it handles a request.
For object mentions in the response to an image and text input, PAS computes an attention-based score showing how strongly the model relied on its own previously generated words.
The closer the PAS score is to zero, the less likely it is that the model has produced a hallucination.
“By understanding the way a vision-language model pays attention to preliminary information, PAS can help identify the exact instance where a model begins to over-rely on its own words,” said Xuan Nhat Hoang, Los Alamos intern.
“Our tool reads signals the AI is already producing, representing a low-overhead way to help ensure that information is reliable and useful,” he added.
PAS could be used in applications where images, documents, diagrams and text are analyzed by vision-language models.
The researchers say it could eventually support reliability checks in areas such as medical imaging, scientific document analysis, engineering diagrams, remote sensing and other visual workflows where unsupported claims could affect downstream decisions.
The Los Alamos team is presenting PAS at the Computer Vision and Pattern Recognition 2026 conference, sponsored by the IEEE and Computer Vision Foundation, in Denver this month.
The work was supported by the Laboratory Directed Research and Development program at Los Alamos.
Get the latest in engineering, tech, space & science - delivered daily to your inbox.
Atharva is a full-time content writer with a post-graduate degree in media & amp; entertainment and a graduate degree in electronics & telecommunications. He has written in the sports and technology domains respectively. In his leisure time, Atharva loves learning about digital marketing and watching soccer matches. His main goal behind joining Interesting Engineering is to learn more about how the recent technological advancements are helping human beings on both societal and individual levels in their daily lives.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。