Picking the right examples makes AI diagnose pathology slides reliably

Vision-language models can reason about medical images, but they're unreliable when shown a few example slides—the choice of examples and wording drastically changes their diagnosis. GAUC selects which examples to show by analyzing the model's internal embedding geometry rather than simple image similarity, combining three optimization goals: matching the full dataset's distribution, handling rephrased questions, and avoiding overconfident guesses. On pathology benchmarks, it improved accuracy and stability over existing selection methods without any fine-tuning.