复旦大学+基于大型语言模型的智能体的兴起与发展-英-86页_5mb

This survey explores the rapid advancement of large language models (LLMs) in creating intelligent AI agents, offering a comprehensive overview of their evolution, design components, applications, and challenges. LLMs now serve as foundational brains modules for agents, enabling traits like autonomy, reactivity, planning, and multimodal perception through enhancements like memory, reasoning, and action modules. Key applications include task-oriented deployments (e.g., web navigation, scientific research), multi-agent systems showcasing cooperation and competition, and human-AI interaction in paradigmes such as instruction-following and equal partnership. Additionally, the concept of "agent society" examines emergent social behaviors, insights from simulations, and ethical risks like adversarial attacks and misuse. The survey discusses mutual benefits between LLM and agent research, evaluation frameworks, security concerns, and path to scaling efficient multi-agent systems. However, it also highlights open problems, including the debated potential of LLMs towards artificial general intelligence and ensuring robust human-AI trust. Overall, it positions LLM-based agents as pivotal to advancing AI, with future work emphasizing safety, scalability, and alignment with human values.