What matters in AI.

Subscribe

An agent workflow cuts mannequin errors in AI images to 8%

A fixed agent workflow let the models abstain when unsure. It cut mannequin errors from a 27% average to 8%.

Researchers used 3 AI image models on 434 images from mobile devices to find military personnel. The models found almost all of them (recall to a maximum of 0.99). Each one identified more than 85% of the mannequin images as military personnel. Two agent workflows, CrewAI and LangGraph, increased precision (the part of finds that is correct) to 0.88 and 0.90.

How it works

The workflows divide the task into steps: the source of the image, the difference between mannequins and persons, and the source of each find. CrewAI uses sub-agents that work together. LangGraph uses a fixed sequence of steps. Each can abstain when confidence is low.

The numbers

The data had 198 military images and 99 mannequin images. LangGraph cut incorrect positives on mannequins to 8% from the 27% average of the models with no workflow. Recall fell a small quantity. LangGraph used 2.8 times more time than one model run.

What is not known

The results come from the tests of the authors. The report does not tell how the workflows work on other types of image. The authors report that the added time is a trade-off.

Insight: A structure around the model fixes the error

In one test, a fixed workflow cut mannequin errors from 27% to 8%, with 2.8 times the time. Nvidia reports that its Switchyard software kept an 80% task success rate for about 33% of the cost. Reliability and cost come from the layer around the model. Get the numbers together. Vendors can possibly sell that layer next.

Sources

  1. Agentic Orchestration Improves Forensic Accuracy in Multimodal AI Image InterpretationBioengineer.org
AI MATTER · NEWS · AI MATTER · NEWS ·19 NOV2026

Posted

Tags

Learn the terms in this story

More in Research

All Research news