← News

Agents & MCPSources: arXiv · 19 Nov 2026

Study finds vision-language model agents lag radiologists when reading emergency CT and MRI scans

A retrospective study reports that six vision-language model agents reading emergency CT and MRI examinations stayed far less accurate than radiologists, often locking in an early hypothesis and rarely correcting it.

In plain words

Vision-language models are AI systems that can look at images and text. Researchers had six of these models read emergency CT and MRI scans under different ways of moving through the images. The models were right far less often than radiologists, and once a model formed an early guess it usually kept it, even when that guess was wrong.

What was announced

The study, published as an arXiv abstract, tested six vision-language model agents on 100 emergency CT and MRI examinations from a public archive, under four navigation conditions: free navigation, navigation starting from a key slice, and all slices or key slices supplied in a single request. Each condition ran twice, for 4800 runs total. Nine radiologists read 30 of the cases, and a large language model judge, validated against two radiologists, graded the answers.

Accuracy was lowest under free navigation at 32.6%, rising to 37.9% with navigation from a key slice, and highest when all slices (42.2%) or key slices (42.5%) were given in one request. On the 30 cases read by both, freely navigating models were correct in 30.0% of runs versus 77.0% for radiologists. Models formed a first hypothesis after a median of 10.8% of slices and kept it in 63.1% of runs; answers were correct in 12.4% of runs after a wrong first hypothesis versus 90.4% after a right one. The authors conclude that across all strategies the agents remained far less accurate than radiologists and seldom corrected a wrong early hypothesis.

What it does not show

  • The evidence is the article abstract only; full methods, case mix and peer review status are not visible in the supplied passage.
  • The study is retrospective and uses cases from a single public archive, so performance in live clinical settings is not established by this evidence.
  • Grading relied on a large language model judge validated against two radiologists, which the abstract does not compare with full radiologist consensus.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.