Dynamic Decision Learning: Test-Time Evolution for Abnormality Grounding in Rare Diseases

Researchers from multiple institutions have introduced Dynamic Decision Learning (DDL), a novel framework that enables frozen large vision-language models (LVLMs) to improve their performance on clinical abnormality grounding, particularly for rare diseases. The approach works by optimizing model instructions and consolidating predictions under visual perturbations at test time, avoiding the need for costly supervised fine-tuning on scarce medical data. This process not only enhances localization quality but also generates a consensus-based reliability score that quantifies model confidence.

Tested on brain imaging benchmarks including a rare-disease dataset covering 281 pathology types, DDL delivered a 105% improvement in mean average precision at IoU 0.75 (mAP@75) for rare disease cases. The framework outperformed both standard adaptation methods and full supervised fine-tuning across models ranging from 3B to 72B parameters. Notably, DDL maintained strong calibration between its reliability scores and actual localization accuracy even under severe distribution shifts and increasing task difficulty, making it a practical solution for real-world clinical deployment where training data is limited.