PortLLM introduces a novel approach that creates lightweight, transferable model patches to personalize LLMs across evolving versions without costly retraining.

• Creates compact portable patches (less than 1MB) that capture domain-specific knowledge
• Enables cross-model portability - patches trained on one LLM can enhance different LLMs
• Demonstrates comparable performance to full fine-tuning while being more resource-efficient
• Facilitates consistent personalization even as base LLMs evolve and update

Medical Impact: PortLLM enables healthcare institutions to maintain specialized medical knowledge in LLMs across model updates without continuous retraining, ensuring consistent performance in sensitive clinical applications while reducing computational costs.

PortLLM: Personalizing Evolving Large Language Models with Training-Free and Portable Model Patches