An LLM wrote a medical image classifier without seeing one image, but 256 labeled scans still beat it on every task

A large language model can write usable weights for a medical image classifier from medical knowledge alone, in one reply, without any labeled images. Such classifiers normally need costly labeled images that are often unavailable, and earlier weight predictors were themselves learned from data; here a general-purpose LLM writes the weights. The networks written by both models were above chance on all six MedMNIST tasks, with mean test AUC of 0.719 for Opus and 0.711 for Astra, and beat zero-shot MedSigLIP on five of them.

The written network has two parts: the first-layer filters form a bank of image features, and the read-out weights state which feature counts as evidence where in the image for each class. The model explains its design in the reply before writing the weights, for example setting filter biases so each filter responds only to a range of grey values and comparing a region with a reference to cancel overall brightness. Evaluation covers seven tasks across chest X-ray, breast ultrasound, abdominal CT in three slice orientations and pelvic MRI, with three independent calls per model that gave broadly consistent results.

The written weights differ systematically from gradient-trained ones: they are sparse and discrete, and the filters resemble textbook operators. The same architecture trained with Adam on 256 labeled images learns smooth, mixed patterns and spreads evidence over the whole image. Opus's networks have 22%–94% of weights exactly zero and use only 8–145 distinct values; on pneumonia and a malignant breast lesion, their evidence maps correlate with Adam's at 0.41 and 0.73.

The demonstrated scope is a single-convolution ProgNet of roughly a thousand parameters on seven tasks: six MedMNIST tasks plus a binary synthetic uterine MRI task, with images and networks small enough to fit one 128k-output reply. Readers should treat it as evidence of feasibility without labels, not as a deployable clinical classifier.

MedMNIST has been public since 2020 and may be in the models' training data; both models did quote class shares that no prompt stated. SynthUterus is recent, so exposure is unlikely but not excluded, its images are synthetic and only 293 in number, and its flexion labels come from 2D slices with some noise. Only two models and one prompt were tested, the prompt did not describe the image layout, and Astra's wrong guess about that layout inverted its networks on that task.