Med-Aligner empowers LLM medical applications for complex medical scenarios

Dear Editor,

Large language models (LLMs) show great promise in medical applications, but challenges like limited high-quality data, closed-source model rigidity, and reasoning degradation during fine-tuning hinder their reliability. To address this, we present Med-Aligner, a plug-in module that learns correction residuals to improve accuracy without full model re-optimization. Trained on 267,524 anonymized medical records from 21 departments, Med-Aligner was integrated with eight LLMs (e.g., GPT-4 and Med-Llama3-8B) and evaluated on helpfulness, harmlessness, and honesty (3H). It achieved average gains of 41.3% ± 25.4%, 30.3% ± 12.4%, and 27.3% ± 14.8% in helpfulness and 10.9% ± 8.6% and 16.6% ± 11.3% in harmlessness and a median 1.7% (range: 0.4%–3.4%) improvement in honesty (p < 0.05). Distribution shift plots confirmed consistent gains, especially in safety and utility. Its lightweight, model-agnostic design enables deployment on resource-limited devices like smartphones. Top rankings on the Alpaca-Eval leaderboard further validate its effectiveness. By bridging open-source and proprietary LLMs, Med-Aligner offers a flexible, efficient solution for medical AI. Limitations include reliance on offline data and the need for clinical validation.

Introduction

LLMs have shown considerable promise in advancing applications within the medical domain.1,2 However, the inherent complexity of medicine requires high levels of accuracy and reliability, which pose significant challenges to the effective application of LLMs.3 First, limited high-quality medical data during pre-training constrain LLM performance in healthcare tasks. Second, closed-source models (e.g., ChatGPT) offer minimal flexibility for parameter tuning, while open-source models (e.g., Llama series), though more adaptable, still face issues like low-quality data, high computational costs, and complex tuning processes. Parameter adjustments may also impair reasoning ability. As LLMs evolve, their outputs can contain subtle errors that are difficult even for experts to detect. RLHF (reinforcement learning with human feedback) -based methods have also reached performance bottlenecks.4 These challenges highlight the need for specialized modules that can augment general-purpose LLMs efficiently.

To address these limitations, we developed Med-Aligner, a plug-in alignment framework for LLMs that enables targeted medical calibration. It improves accuracy and reliability in domain-specific tasks by combining the strengths of open- and closed-source models without requiring full re-optimization.