將大型語言模型（LLMs）適應於醫療領域對於臨床決策支援至關重要。然而，傳統持續預訓練（CPT）面臨高昂計算成本與災難性遺忘，後者對必須處理多樣指令的臨床安全構成威脅。為此，我們提出一個高效的領域適應框架，包含兩項核心貢獻：(1) 對比式梯度敏感度（Contrastive Gradient Sensitivity， CGS），透過比較醫療與通用語料的梯度響應，自動更新最具領域鑑別力的模組，以在保護通用能力的同時最大化醫療知識吸收；(2) 困惑度導向的資料純度過濾，以精煉訓練語料。我們在 Llama-3.1-8B、Qwen2.5-7B 與 Mistral-7B-v0.3 上進行評估。在僅更新 8% 參數的條件下，CGS 在 MedQA 上均達到與 LISA 基線相當或略優的表現。在 Llama-3.1-8B 上，CGS 取得 63.50% 的 MedQA 準確率，達到與全參數微調的 Meditron3-8B（63.39%）相當的水準，且 MMLU 通用能力幾乎無損（68.47% vs. 原始 68.54%）。自動評估框架與人類一致性達 Cohen’s $\kappa = 1.00$ ($n = 200$)。消融實驗支持了 CGS 定向更新與資料過濾在平衡領域適應與臨床安全上的貢獻。;Adapting large language models (LLMs) to the medical domain is crucial for clinical decision support. However， standard continual pre-training (CPT) to inject clinical knowledge often suffers from prohibitive computational costs and catastrophic forgetting of general capabilities—a critical safety risk for clinical LLMs that must comprehend diverse， general instructions. To address this， we propose an efficient medical domain adaptation framework with two core contributions:(1) Contrastive Gradient Sensitivity (CGS)， a gradient-informed parameter selection strategy that compares gradient responses across medical and general corpora to identify and update the most domain-discriminative modules， protecting general capabilities while maximizing medical knowledge absorption; and(2) Perplexity-guided data purity filtering to refine the training corpus.We evaluate our framework on Llama-3.1-8B， Qwen2.5-7B， and Mistral-7B-v0.3 across major medical benchmarks (MedQA， MedMCQA， PubMedQA) and general tasks (MMLU). Updating only 8% of parameters， CGS achieves comparable or slightly higher MedQA accuracy than the LISA baseline across all models. On Llama-3.1-8B， CGS achieves 63.50% on MedQA—achieving comparable performance to the fully pre-trained Meditron3-8B (63.39%)—while preserving near-intact general capabilities on MMLU (68.47% vs. 68.54% base). Our automatic evaluation pipeline is validated by a human agreement of Cohen’s $\kappa = 1.00$ ($n = 200$). Ablation studies support the importance of CGS’s targeted parameter updates and perplexity-based filtering in balancing domain knowledge absorption and safety-critical capability preservation.
