大型語言模型規模持續擴大，模型壓縮已成為實際部署的關鍵需求；然而壓縮方法繁多，且在精度、延遲、記憶體與模型大小之間存在複雜權衡，對開發者構成技術門檻。 本研究提出名為 LLM-as-Strategist 的多目標壓縮框架，以 LLM 為協調器：開發者僅需提供模型名稱與資料集，框架即整合量化、剪枝、低秩近似與微調等方法，於精度、延遲、記憶體、大小與碳排放五個維度上自動搜索 Pareto 最優解，並以超體積衡量前沿品質。 在 GPT-2 與 Qwen2.5 系列上的系統性比較顯示，在有限方法下，LLM 協調、自動搜索與隨機策略收斂至相近的 Pareto 前沿，呈現搜索空間飽和；量化的精度保留明顯優於剪枝與低秩近似；GPT-4o-mini 與 GPT-4o 表現相當，但成本僅約十分之一，為最具成本效益的協調器。此外，本地中小型模型受限於結構化輸出能力而難以勝任協調角色，瓶頸在於模型規模而非壓縮狀態。 本研究將模型壓縮重新框架為可自動探索、可公平比較的多目標系統工程問題，並將碳排放納入目標維度以回應 Green AI 的效率訴求。;As large language models continue to scale， model compression has become essential for practical deployment. Yet the many compression methods， and their complex trade-offs among accuracy， latency， memory， and model size， pose a technical barrier for developers. This study proposes LLM-as-Strategist， a multi-objective compression framework with an LLM as the coordinator. Given only a model name and a dataset， the framework integrates quantization， pruning， low-rank approximation， and fine-tuning， and automatically searches for Pareto-optimal solutions across accuracy， latency， memory， size， and carbon emissions， with frontier quality measured by Hypervolume. A systematic comparison on GPT-2 and the Qwen2.5 series shows that， under a limited method library， LLM coordination， automated search， and random selection converge to similar Pareto frontiers， indicating search-space saturation; that quantization retains accuracy markedly better than pruning or low-rank approximation; and that GPT-4o-mini matches GPT-4o at roughly one-tenth the cost， making it the most cost-effective coordinator. Moreover， small local models are inadequate as coordinators because of their limited structured-output capability， with the bottleneck being model scale rather than compression state. The value of this study lies in reframing model compression as a multi-objective systemsengineering problem that can be automatically explored and fairly compared， and in incorporating carbon emissions as an objective dimension in response to the efficiency goals of Green AI
