Lost in the Hype: Revealing and Dissecting the Performance Degradation of Medical Multimodal Large Language Models in Image Classification
A systematic teardown of 14 medical AI models finds they consistently fail at basic image classification — exposing four specific reasons why clinical promises remain unmet.
The Thesis
Multimodal large language models (MLLMs — AI systems that process both images and text) have been aggressively marketed for medical imaging, but this paper shows they routinely lose to older, simpler deep learning classifiers on basic diagnostic tasks. The researchers didn't just measure the gap; they traced exactly where image information degrades as it flows through each layer of the model pipeline. They identified four distinct failure points: poor visual feature quality, information lost when translating image signals into the language model's format, weak reasoning about medical image content, and a mismatch between visual patterns and clinical meaning. The practical implication is stark — hospitals and health systems betting on off-the-shelf medical MLLMs for image-based diagnosis may be deploying systems that underperform tools they already own.
Catalyst
Medical MLLMs have proliferated rapidly since 2023, with dozens of open-source and commercial variants claiming clinical-grade performance. This paper is possible now because enough named, comparable models exist to run a structured 14-model comparison. The field has also developed 'feature probing' techniques — diagnostic tools borrowed from interpretability research that can interrogate a neural network layer by layer — mature enough to apply systematically across an entire model class.
What's New
Prior benchmarking work in medical AI mostly compared final accuracy scores across models, treating each system as a black box. That approach identifies that a gap exists but not why. This paper applies feature probing — a method where small classifier heads are trained on internal model representations at each layer to measure how much diagnostic signal survives at each processing stage. The result is the first explicit, module-by-module map of where medical image classification information is degraded or destroyed inside an MLLM.
The Counter
This paper benchmarks 14 open-source models on three classification datasets — a reasonable sample, but classification is arguably the task MLLMs are least designed to optimize for. Most clinical MLLM deployments target report generation, anomaly description, or clinical reasoning assistance, not pure label prediction. Judging these systems on classification accuracy is a bit like grading a radiologist solely on how fast they sort X-rays into bins. Additionally, the feature probing methodology, while novel in this context, introduces its own assumptions: a small classifier head that fails to extract signal from a layer doesn't definitively prove signal is absent, only that it isn't linearly separable. The paper also focuses entirely on open-source models, which may systematically lag commercial offerings from large health system AI vendors. Finally, the four failure modes are descriptive categories derived from one analytical framework — they aren't independently validated, and reasonable researchers might carve the same evidence differently.
Longs
- ISRG (Intuitive Surgical) — incumbents in validated, cleared medical imaging pipelines benefit as MLLM hype cools
- GEHC (GE HealthCare) — established medical imaging AI with regulatory track record looks stronger relative to unproven MLLMs
- VEEV (Veeva Systems) — clinical data infrastructure plays benefit if MLLMs need better training pipelines
- RXRX (Recursion Pharmaceuticals) — biotech using purpose-built vision models rather than general MLLMs
- NVDA — indirectly, as diagnostic compute shifts back toward specialized vision model training rather than LLM inference
Shorts
- Startups marketing general-purpose medical MLLMs for diagnostic classification without rigorous validation — the paper directly undermines their core sales claim
- Health systems that purchased or are piloting MLLM-based diagnostic tools under the assumption they match or exceed traditional CNN (convolutional neural network) classifiers
- Venture-backed companies that raised on MLLM medical imaging narratives without task-specific benchmarking — investor scrutiny will increase
Enablers (Picks & Shovels)
- Hugging Face medical model hub — the open-source ecosystem from which the 14 tested MLLMs were drawn
- MIMIC and similar open clinical imaging datasets — required for the classification benchmarks used in this study
- Feature probing toolkits (e.g., diagnostic classifier libraries in PyTorch) — the core interpretability method the paper deploys
- MedImageNet and PathMNIST — benchmark datasets representative of those used in the paper's experiments
Private Watchlist
- Rad AI — radiology-specific AI using task-tuned models rather than general MLLMs
- Enlitic — medical imaging company focused on structured, validated classification pipelines
- Segmed — medical imaging data curation, addresses training data quality gaps identified in the paper
- Gradient Health — curated medical imaging datasets for model training
Resources
The Paper
The rise of multimodal large language models (MLLMs) has sparked an unprecedented wave of applications in the field of medical imaging analysis. However, as one of the earliest and most fundamental tasks integrated into this paradigm, medical image classification reveals a sobering reality: state-of-the-art medical MLLMs consistently underperform compared to traditional deep learning models, despite their overwhelming advantages in pre-training data and model parameters. This paradox prompts a critical rethinking: where exactly does the performance degradation originate? In this paper, we conduct extensive experiments on 14 open-source medical MLLMs across three representative image classification datasets. Moving beyond superficial performance benchmarking, we employ feature probing to track the information flow of visual features module-by-module and layer-by-layer throughout the entire MLLM pipeline, enabling explicit visualization of where and how classification signals are distorted, diluted, or overridden. As the first attempt to dissect classification performance degradation in medical MLLMs, our findings reveal four failure modes: 1) quality limitation in visual representation, 2) fidelity loss in connector projection, 3) comprehension deficit in LLM reasoning, and 4) misalignment of semantic mapping. Meanwhile, we introduce quantitative scores that characterize the healthiness of feature evolution, enabling principled comparisons across diverse MLLMs and datasets. Furthermore, we provide insightful discussions centered on the critical barriers that prevent current medical MLLMs from fulfilling their promised clinical potential. We hope that our work provokes rethinking within the community-highlighting that the road from high expectations to clinically deployable MLLMs remains long and winding.