Medical VLMs show confidence despite contradicting their own reasoning
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
A new evaluation framework exposes a critical reliability gap in medical vision-language models: their confidence scores and written explanations fail to track actual performance degradation. Researchers tested four VLMs on mammography tasks, finding that models maintain high confidence even when images are perturbed, and that option ordering shifts answers without corresponding confidence adjustments. Most concerning, models sometimes contradict their own stated reasoning in final predictions. This work signals that current VLM evaluation metrics miss internal incoherence, a problem that matters acutely in high-stakes domains like radiology where users rely on model transparency as a proxy for trustworthiness.
Modelwire’s AI-generated summary of coverage from arXiv cs.CL.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
The paper isolates a failure mode that standard accuracy metrics miss: models can maintain high confidence while their reasoning contradicts their predictions, and confidence remains stable even when image perturbations degrade actual performance. This is not just miscalibration; it's incoherence between what models claim to know and what they actually output.
This work extends a pattern visible in recent VLM evaluation research. The crochet paper (early November) showed that plausible-sounding outputs fail under real-world execution; this mammography study reveals the inverse problem: models sound confident while internally contradicting themselves. Both expose gaps between benchmark metrics and actual reliability. More directly, the linear representations paper from the same day identifies that language models encode uncertainty signals in their activations but fail to surface them in confidence outputs. Here we see that failure mode weaponized in medical imaging, where users treat confidence as a proxy for trustworthiness.
If the same four VLMs are re-evaluated on BI-RADS classification using the authors' coherence framework and show that retraining on confidence-reasoning alignment improves both calibration and clinical decision support in a prospective radiology study within 12 months, that confirms this is a solvable architectural problem rather than an inherent VLM limitation.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.CL
Language models encode ambiguity but fail to report it reliably
Researchers have identified a structural gap in how language models assess their own reliability. The work reveals that LMs encode representations of input ambiguity within their activation patterns, yet fail to leverage this signal for confidence calibration. By extracting these latent ambiguity signals through contrastive analysis and proposing DART, a targeted fine-tuning approach, the team…
Mentionsvision-language models · mammography · BI-RADS · breast density classification
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Evaluating Cross-Signal Coherence in Mammography Vision-Language Models”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.