Researchers found a shortcut for teaching AI models to read medical images: skip the retraining, just show them better examples.
The technique, called representation-guided in-context learning (RG-ICL), lets a general-purpose multimodal AI model improve at reading medical images without any fine-tuning. Instead of retraining the model on medical data, RG-ICL uses separate, already-trained encoders to find past cases that closely match a new query image, then feeds those as reference examples at the moment of inference. Tested across eight datasets covering histopathology, radiology, and retinal scans, it lifted classification accuracy by an average of 20 percentage points and question-answering accuracy by 13 points over giving the model no examples or random ones, closing in on or beating models that were actually fine-tuned for the job.
The more interesting finding is about quality over quantity: six well-matched reference cases beat 32 randomly chosen ones, and mismatched examples could drag accuracy below a model with no examples at all. That is a useful data point in the broader argument over whether AI progress in specialized fields requires expensive retraining or just smarter retrieval. It also matters for hospitals and clinics, which often lack the labeled data or compute budget to fine-tune models but do have libraries of past cases to draw from.
Still, this is an arXiv preprint, not a peer-reviewed clinical tool, and picking six perfect examples in a lab benchmark is a different problem than doing it fast and reliably on messy real-world scans.