LLM 驱动的自主智能体

LLM Powered Autonomous Agents

翁荔 Lilian Weng · OpenAI · 2023-06-23 · Lil'Log ↗

摘要 · Abstract

日期：2023 年 6 月 23 日 | 预计阅读时间：31 分钟 | 作者：Lilian Weng 以大型语言模型（LLM）作为核心控制器构建智能体是一个很酷的概念。一些概念验证演示，如 AutoGPT、GPT-Engineer 和 BabyAGI，是鼓舞人心的例子。LLM 的潜力不仅限于生成写得好的文案、故事、论文和程序；它可以被框架化为一个强大的通用问题解决器。

Date: June 23, 2023 | Estimated Reading Time: 31 min | Author: Lilian Weng Building agents with LLM (large language model) as its core controller is a cool concept. Several proof-of-concepts demos, such as AutoGPT, GPT-Engineer and BabyAGI, serve as inspiring examples. The potentiality of LLM extends beyond generating well-written copies, stories, essays and programs; it can be framed as a powerful general problem solver.

核心贡献 · Key contributions

提出了一个以 LLM 为核心的自主智能体综合框架，包含规划、记忆和工具使用。
Proposes a comprehensive framework for LLM-powered autonomous agents with planning, memory, and tool use.
引入了自我反思机制（如 Reflexion 和 Chain of Hindsight）以实现迭代改进。
Introduces self-reflection mechanisms like Reflexion and Chain of Hindsight for iterative improvement.
展示了结合短期（上下文内）和长期（向量存储）记忆的记忆架构。
Demonstrates memory architectures combining short-term (in-context) and long-term (vector store) memory.
展示了通过微调（Toolformer）和任务规划（HuggingGPT）使用工具来扩展 LLM 能力。
Shows tool use via fine-tuning (Toolformer) and task planning (HuggingGPT) to extend LLM capabilities.
提出了在沙盒环境中具有涌现社会行为的生成式智能体。
Presents generative agents with emergent social behaviors in sandbox environments.
回顾了 AutoGPT 和 GPT-Engineer 等实际演示，突出了潜力和可靠性问题。
Reviews practical demos like AutoGPT and GPT-Engineer, highlighting potential and reliability issues.

局限 · Limitations

有限的上下文窗口限制了历史信息和详细指令，制约了长期记忆。
Finite context window restricts historical information and detailed instructions, limiting long-term memory.
长期规划和任务分解仍具挑战性；LLM 难以应对意外错误。
Long-term planning and task decomposition remain challenging; LLMs struggle with unexpected errors.
自然语言接口不可靠；LLM 会产生格式错误和偶尔的叛逆行为。
Natural language interface is unreliable; LLMs produce formatting errors and occasional rebellious behavior.
工具使用依赖于 LLM 正确选择和调用 API 的能力，这并非总是可靠。
Tool use depends on LLM's ability to correctly select and invoke APIs, which is not always reliable.
自我反思和记忆检索可能无法完全弥补复杂任务中有限的注意力跨度。
Self-reflection and memory retrieval may not fully compensate for limited attention span in complex tasks.

论文章节 · Sections（共 14）

概述 Overview
智能体系统概述 Agent System Overview#(https://lilianweng.github.io/posts/2023-06-23-agent/#agent-system-overview)
组件一：规划 Component One: Planning#(https://lilianweng.github.io/posts/2023-06-23-agent/#component-one-planning)
任务分解 Task Decomposition#(https://lilianweng.github.io/posts/2023-06-23-agent/#task-decomposition)
自我反思 Self-Reflection#(https://lilianweng.github.io/posts/2023-06-23-agent/#self-reflection)
组件二：记忆 Component Two: Memory#(https://lilianweng.github.io/posts/2023-06-23-agent/#component-two-memory)
记忆类型 Types of Memory#(https://lilianweng.github.io/posts/2023-06-23-agent/#types-of-memory)
最大内积搜索（MIPS） Maximum Inner Product Search (MIPS)#(https://lilianweng.github.io/posts/2023-06-23-agent/#maximum-inner-product-search-mips)
组件三：工具使用 Component Three: Tool Use#(https://lilianweng.github.io/posts/2023-06-23-agent/#component-three-tool-use)
科学发现智能体 Scientific Discovery Agent#(https://lilianweng.github.io/posts/2023-06-23-agent/#scientific-discovery-agent)
生成式智能体模拟 Generative Agents Simulation#(https://lilianweng.github.io/posts/2023-06-23-agent/#generative-agents-simulation)
概念验证示例 Proof-of-Concept Examples#(https://lilianweng.github.io/posts/2023-06-23-agent/#proof-of-concept-examples)
假设与实现步骤 … same conversation as earlier, ended with "Make your own assumptions and state them explicitly before starting".
挑战 Challenges#(https://lilianweng.github.io/posts/2023-06-23-agent/#challenges)

阅读逐段中英对照全文 →