The paper discusses the evaluation of Large Language Models (LLMs) to determine their capabilities and suitability for various tasks. It proposes a two-stage framework: from "core ability" to "agent," explaining how LLMs can be applied based on their specific capabilities and the evaluation methods in each stage. The core ability stage focuses on reasoning, societal impact, and domain knowledge, while the agent stage covers embodied action, planning, and tool learning. The paper also examines the challenges in evaluating LLMs and suggests future directions for development. The contributions include a detailed framework for evaluating LLMs, an analysis of current performance levels, and insights into the broader societal implications of LLMs. The paper reviews existing evaluation methods and datasets, providing a comprehensive overview of the state-of-the-art in LLM evaluation.The paper discusses the evaluation of Large Language Models (LLMs) to determine their capabilities and suitability for various tasks. It proposes a two-stage framework: from "core ability" to "agent," explaining how LLMs can be applied based on their specific capabilities and the evaluation methods in each stage. The core ability stage focuses on reasoning, societal impact, and domain knowledge, while the agent stage covers embodied action, planning, and tool learning. The paper also examines the challenges in evaluating LLMs and suggests future directions for development. The contributions include a detailed framework for evaluating LLMs, an analysis of current performance levels, and insights into the broader societal implications of LLMs. The paper reviews existing evaluation methods and datasets, providing a comprehensive overview of the state-of-the-art in LLM evaluation.