Frozen Models, Evolving Expertise: Model-Agnostic Learning from Deployment Experience for Multimodal Medical AI
The paper proposes a way for deployed medical LLMs and VLMs to improve without changing their weights.
The framework stores deployment experience outside the model as reasoning skills, vetted factual memory, and visual case examples. It accepts an update only when it helps on new cases without hurting earlier performance. The authors report gains across six benchmarks and four base models, including up to 34.2% over the base model on medical tasks. HF Daily Papers' note
The framework stores deployment experience outside the model as reasoning skills, vetted factual memory, and visual case examples. It accepts an update only when it helps on new cases without hurting earlier performance. The authors report gains across six benchmarks and four base models, including up to 34.2% over the base model on medical tasks. HF Daily Papers' note
score 4