# New framework lets frozen medical AI models keep learning after deployment

> Researchers propose a model‑agnostic system that updates frozen LLMs and VLMs with real‑world cases, boosting performance up to 34% on medical tasks.

Oossa · 2026-10-08 · https://oossa.com/en/new-framework-lets-frozen-medical-ai-models-keep-learning-after-deployment

A team led by Yexiao He released a paper on Oct 8, 2026 describing a framework that lets frozen large language models (LLMs) and vision‑language models (VLMs) learn from the cases they handle. The system adds three types of external expertise – a Skill for reasoning, a Knowledge Memory for facts, and a Multimodal Knowledge Base for visual examples. It updates the model only when new cases improve performance without hurting earlier results. Tests on six medical benchmarks with four different base models showed gains of up to 34.2%.

## The facts

- Improvement of up to 34.2% on medical tasks
- Evaluated on six benchmarks using four base models

## Why it matters

Clinicians could rely on AI tools that stay current with new guidelines without needing costly retraining.

## Sources & references

1. [Frozen Models, Evolving Expertise: Model-Agnostic Learning from Deployment Experience for Multimodal Medical AI](https://arxiv.org/abs/2610.09146) – arXiv, 2026-10-08

Last updated: 2026-10-08