MedAdapter is a novel test-time adaptation method designed to enhance the biomedical reasoning capabilities of large language models (LLMs). It addresses the challenges of adapting LLMs to the biomedical domain, which are exacerbated by their large size and corporate privacy concerns. MedAdapter introduces a small BERT-sized adapter that fine-tunes only a subset of parameters to rank candidate solutions generated by the LLMs, effectively adapting the original model without requiring extensive computational resources or data sharing. The method is evaluated on five biomedical QA datasets, demonstrating significant performance improvements. For white-box LLMs, MedAdapter achieves 99.35% of supervised fine-tuning performance using only 14.75% of the GPU memory. For black-box LLMs, it achieves comparable or superior performance to fine-tuning via APIs, with a 11.31% average improvement at only 15.59% of the cost. MedAdapter also outperforms other adaptation methods when combined with train-time adaptation, highlighting its flexibility and complementarity. MedAdapter provides a practical, privacy-preserving, cost-effective, and transparent solution for adapting LLMs to biomedical applications, balancing model performance, computational resources, and data privacy.MedAdapter is a novel test-time adaptation method designed to enhance the biomedical reasoning capabilities of large language models (LLMs). It addresses the challenges of adapting LLMs to the biomedical domain, which are exacerbated by their large size and corporate privacy concerns. MedAdapter introduces a small BERT-sized adapter that fine-tunes only a subset of parameters to rank candidate solutions generated by the LLMs, effectively adapting the original model without requiring extensive computational resources or data sharing. The method is evaluated on five biomedical QA datasets, demonstrating significant performance improvements. For white-box LLMs, MedAdapter achieves 99.35% of supervised fine-tuning performance using only 14.75% of the GPU memory. For black-box LLMs, it achieves comparable or superior performance to fine-tuning via APIs, with a 11.31% average improvement at only 15.59% of the cost. MedAdapter also outperforms other adaptation methods when combined with train-time adaptation, highlighting its flexibility and complementarity. MedAdapter provides a practical, privacy-preserving, cost-effective, and transparent solution for adapting LLMs to biomedical applications, balancing model performance, computational resources, and data privacy.