Generative Agents: Interactive Simulacra of Human Behavior
生成代理：人类行为的交互模拟体

Interactive artificial intelligence systems aim to combine human insights and capabilities in computational artifacts that can augment their users  (Amershi et al., 2014; Fails and Olsen Jr, 2003). A long line of work has explored ways to enable users to interactively specify model behavior. For instance, Crayons demonstrated an early vision of interactive machine learning, allowing non-expert users to train classifiers (Fails and Olsen Jr, 2003). Further work helped to articulate how end users might describe their classification goals to the system through examples (Fogarty et al., 2008) or demonstration (Fiebrink and Cook, 2010). Recent advancements have extended these explorations to deep learning (Lam et al., 2023) and prompt-based authoring (Jiang et al., 2022; Wu et al., 2022b; Liu et al., 2022).
交互式人工智能系统旨在将人类的见解和能力与计算工件相结合，从而增强其用户（Amershi 等，2014 年；Fails 和 Olsen Jr，2003 年）。长期以来，人们一直在探索使用户能够交互地指定模型行为的方法。例如，Crayons 展示了交互式机器学习的早期愿景，允许非专家用户训练分类器（Fails 和 Olsen Jr，2003 年）。进一步的工作有助于阐明最终用户如何通过示例（Fogarty 等，2008 年）或演示（Fiebrink 和 Cook，2010 年）描述其分类目标给系统。最近的进展将这些探索扩展到深度学习（Lam 等，2023 年）和基于提示的创作（Jiang 等，2022 年；Wu 等，2022b；Liu 等，2022 年）。

Meanwhile, a persistent thread of research has advanced the case for language- and agent-based interaction in human-computer interaction. Formative work such as SHRDLU (Winograd, 1971) and ELIZA (Weizenbaum, 1966) demonstrated the opportunities and the risks associated with natural language interaction with computing systems. As research progressed, it became evident that autonomous agents could offer new metaphors for delegation and interaction (Maes, 1995), but the boundaries of delegation between humans and agents have remained the subject of ongoing debate and refinement (Shneiderman and Maes, 1997; Horvitz, 1999; Shneiderman, 2022). Recently, this technology has reached a level of stability that enables agents to interact via natural language in large and complex online social environments (e.g., (Krishna et al., 2022)). Natural language interaction provides a novel modality that can enhance user abilities in domains such as photo editing (Linder et al., 2013; Fourney et al., 2011; Adar et al., 2014) and code editing (Rong et al., 2016).
与此同时，一系列持续的研究已经推动了语言和基于代理的人机交互的案例。形成性工作，如 SHRDLU（Winograd，1971 年）和 ELIZA（Weizenbaum，1966 年）展示了与计算系统进行自然语言交互相关的机会和风险。随着研究的进展，人们逐渐意识到自主代理可以为委托和交互提供新的隐喻（Maes，1995 年），但人类和代理之间的委托边界仍然是持续辩论和完善的主题（Shneiderman 和 Maes，1997 年；Horvitz，1999 年；Shneiderman，2022 年）。最近，这项技术已经达到了一种稳定水平，使代理能够通过自然语言在大型和复杂的在线社交环境中进行交互（例如，（Krishna 等，2022 年））。自然语言交互提供了一种新颖的模式，可以增强用户在诸如照片编辑（Linder 等，2013 年；Fourney 等，2011 年；Adar 等，2014 年）和代码编辑（Rong 等，2016 年）等领域的能力。

We convene these threads of work to show that we can now create agents that proxy human behavior for interactive systems, and interact with them using natural language. In doing so, this work reopens the door to examining foundational human-computer interaction questions around cognitive models such as GOMS and Keystroke-Level Model (KLM) (Card et al., 1983, 1980), around prototyping tools (Park et al., 2022), and around ubiquitous computing applications (Weiser, 1991; Dey, 2001; Fast et al., 2016).
我们召开这些工作线索，以展示我们现在可以为交互系统创建代理人，代理人可以模拟人类行为，并使用自然语言与它们互动。这样做，这项工作重新打开了研究基础人机交互问题的大门，如 GOMS 和 Keystroke-Level Model（KLM）（Card 等，1983 年，1980 年），原型工具（Park 等，2022 年）以及普遍计算应用（Weiser，1991 年；Dey，2001 年；Fast 等，2016 年）。

Prior literature has described believability, or believable agents, as a central design and engineering goal. Believable agents are designed to provide an illusion of life and present a facade of realism in the way they appear to make decisions and act on their own volition, similar to the characters in Disney movies (Bates, 1994; Thomas and Johnston, 1981). These agents can populate and perceive an open world environment like the one we inhabit (Laird and VanLent, 2001; Bates, 1994), and strive to behave in ways that exhibit emergent behaviors grounded in social interactions with users or other agents with the aim of becoming believable proxies of our behavior in hypothetical simulations of individuals and communities (McCoy et al., 2012; Burkinshaw, 2009; Francis, 2010). Historically, these agents were developed in the context of intelligent game non-player characters (NPCs) (Laird and VanLent, 2001; Riedl, 2012). Creating NPCs with believable behavior, if possible, could enhance player experiences in games and interactive fictions by enabling emergent narratives (Swartout et al., 2006; Aylett, 1999; Brenner, 2010; Ibister and Nass, 2000) and social interactions with the agents (Zubek, 2002). However, more importantly, game worlds provide increasingly realistic representations of real-world affordances, and as observed by Laird and van Lent in 2001, these simulated worlds offer accessible testbeds for developers of believable agents to finesse the agents’ cognitive capabilities without worrying about implementing robotics in the real world or creating simulation environments from scratch (Laird and VanLent, 2001; Riedl, 2012).
先前的文献已经描述了可信度，或者说可信的代理，作为一个中心设计和工程目标。可信的代理被设计为提供一种生命的幻觉，并以一种看似自主地做出决策和行动的现实幻觉出现，类似于迪士尼电影中的角色（Bates，1994 年；Thomas 和 Johnston，1981 年）。这些代理可以在我们所居住的开放世界环境中生存和感知（Laird 和 VanLent，2001 年；Bates，1994 年），并努力以展现出社交互动中的新兴行为为基础的方式行为，以成为我们在个体和社区的假设模拟中行为的可信代理的目标（McCoy 等人，2012 年；Burkinshaw，2009 年；Francis，2010 年）。从历史上看，这些代理是在智能游戏非玩家角色（NPCs）的背景下开发的（Laird 和 VanLent，2001 年；Riedl，2012 年）。如果可能的话，创造出具有可信行为的 NPCs 可以通过启用新兴叙事来增强玩家在游戏和互动虚构中的体验（Swartout 等人）。, 2006; Aylett, 1999; Brenner, 2010; Ibister and Nass, 2000) and social interactions with the agents (Zubek, 2002). However, more importantly, game worlds provide increasingly realistic representations of real-world affordances, and as observed by Laird and van Lent in 2001, these simulated worlds offer accessible testbeds for developers of believable agents to finesse the agents’ cognitive capabilities without worrying about implementing robotics in the real world or creating simulation environments from scratch (Laird and VanLent, 2001; Riedl, 2012).

A diverse set of approaches to creating believable agents emerged over the past four decades. In implementation, however, these approaches often simplified the environment or dimensions of agent behavior to make the effort more manageable (Brooks et al., 2000; Minsky and Papert, 1970). Rule-based approaches, such as finite-state machines (Siu et al., 2021; Umarov et al., 2012) and behavior trees (Knafla, 2011; Pillosu, 2009; Hecker, 2011) account for the brute force approach of human-authoring the agent’s behavior (McCoy et al., 2012). They provide a straightforward way of creating simple agents that is still the most dominant approach today (McCoy et al., 2009; Miyashita et al., 2017; Yannakakis, 2012), and can even handle rudimentary social interactions, as shown in games such as Mass Effect (BioWare, 2007) and The Sims (Arts, 2009) series. Nonetheless, manually crafting behavior that can comprehensively address the breadth of possible interactions in an open world is untenable. This means that the resulting agent behaviors may not fully represent the consequences of their interactions (McCoy et al., 2012, 2011a, 2011b), and cannot perform new procedures that were not hard-coded in their script (Siu et al., 2021; Umarov et al., 2012). On the other hand, prevalent learning-based approaches for creating believable agents, such as reinforcement learning, have overcome the challenge of manual authoring by letting the agents learn their behavior, and have achieved superhuman performance in recent years in games such as AlphaStar for Starcraft (Vinyals et al., 2019) and OpenAI Five for Dota 2 (Berner et al., 2019). However, their success has largely taken place in adversarial games with readily definable rewards that a learning algorithm can optimize for. They have not yet addressed the challenge of creating believable agents in an open world (Siu et al., 2021; Miyashita et al., 2017; Hausknecht et al., 2020).
一系列不同的方法出现在过去的四十年中，用于创建可信的代理。然而，在实施过程中，这些方法通常简化了环境或代理行为的维度，以使工作更易管理。基于规则的方法，如有限状态机和行为树，解释了人类编写代理行为的蛮力方法。它们提供了一种简单的方式来创建简单的代理，这仍然是今天最主流的方法，甚至可以处理基本的社交互动，如《质量效应》和《模拟人生》系列游戏中所示。然而，手动制作可以全面解决开放世界中可能互动的广泛范围的行为是不可行的。这意味着最终的代理行为可能无法完全代表它们互动的后果。, 2012, 2011a, 2011b), and cannot perform new procedures that were not hard-coded in their script (Siu et al., 2021; Umarov et al., 2012). On the other hand, prevalent learning-based approaches for creating believable agents, such as reinforcement learning, have overcome the challenge of manual authoring by letting the agents learn their behavior, and have achieved superhuman performance in recent years in games such as AlphaStar for Starcraft (Vinyals et al., 2019) and OpenAI Five for Dota 2 (Berner et al., 2019). However, their success has largely taken place in adversarial games with readily definable rewards that a learning algorithm can optimize for. They have not yet addressed the challenge of creating believable agents in an open world (Siu et al., 2021; Miyashita et al., 2017; Hausknecht et al., 2020).

Cognitive architectures in computation, pioneered by Newell, aimed to build the infrastructure for supporting a comprehensive set of cognitive functions (Newell, 1990) that suited the all-encompassing nature of believable agents held in its original vision. They fueled some of the earliest examples of believable agents. For instance, Quakebot-SOAR (Laird, 2000) and ICARUS (Langley et al., 2005; Choi et al., 2021) generated NPCs in first-person shooter games, while TacAir-SOAR (Pew and Mavor, 1998) generated pilots in aerial combat training simulations. The architectures used by these agents differed (Quakebot- and TacAir-SOAR relied on SOAR (Laird, 2012), while ICARUS relied on its own variation that was inspired by SOAR and ACT-R (Anderson, 1993)), but they shared the same underlying principle (Laird et al., 2017). They maintained short-term and long-term memories, filled these memories with symbolic structures, and operated in perceive-plan-act cycles, dynamically perceiving the environment and matching it with one of the manually crafted action procedures (Umarov et al., 2012; Laird, 2001). Agents created using cognitive architectures aimed to be generalizable to most, if not all, open world contexts and exhibited robust behavior for their time. However, their space of action was limited to manually crafted procedural knowledge, and they did not offer a mechanism through which the agents could be inspired to seek new behavior. As such, these agents were deployed mostly in non-open world contexts such as first-person shooter games (Laird, 2000; Choi et al., 2021) or blocks worlds (Langley et al., 2005).
认知架构在计算中的先驱者纽厄尔旨在构建支持全面认知功能集合的基础设施（纽厄尔，1990），以适应其最初愿景中所持有的全方位可信代理的性质。它们推动了一些最早的可信代理的例子。例如，Quakebot-SOAR（Laird，2000）和 ICARUS（Langley 等，2005；Choi 等，2021）在第一人称射击游戏中生成 NPC，而 TacAir-SOAR（Pew 和 Mavor，1998）在空中战斗训练模拟中生成飞行员。这些代理使用的架构不同（Quakebot-和 TacAir-SOAR 依赖于 SOAR（Laird，2012），而 ICARUS 则依赖于受 SOAR 和 ACT-R（Anderson，1993）启发的自身变体），但它们共享相同的基本原则（Laird 等，2017）。它们保持短期和长期记忆，用符号结构填充这些记忆，并在感知-规划-行动循环中运行，动态感知环境并将其与手工制作的行动程序之一匹配（Umarov 等，2012；Laird，2001）。 使用认知架构创建的代理人旨在通用化到大多数，如果不是全部，开放世界环境，并展示了他们所处时代的强大行为。然而，他们的行动空间仅限于手工制作的程序化知识，并没有提供一种机制，通过该机制代理人可以被激发寻求新行为。因此，这些代理人主要部署在非开放世界环境中，如第一人称射击游戏（Laird，2000；Choi 等，2021）或方块世界（Langley 等，2005）。

Today, creating believable agents as described in its original definition remains an open problem (Yannakakis, 2012; Riedl, 2012). Many have moved on, arguing that although current approaches for creating believable agents might be cumbersome and limited, they are good enough to support existing gameplay and interactions (Yannakakis, 2012; Champandard, 2012; Nareyek, 2007). Our argument is that large language models offer an opportunity to re-examine these questions, provided that we can craft an effective architecture to synthesize memories into believable behavior. We offer a step toward such an architecture in this paper.
今天，根据其原始定义描述的创建可信代理仍然是一个悬而未决的问题（Yannakakis，2012 年；Riedl，2012 年）。许多人已经前进，认为尽管当前用于创建可信代理的方法可能繁琐且有限，但它们已足以支持现有的游戏玩法和互动（Yannakakis，2012 年；Champandard，2012 年；Nareyek，2007 年）。我们的论点是，大型语言模型提供了重新审视这些问题的机会，前提是我们能够构建一个有效的架构，将记忆合成为可信行为。我们在本文中提供了朝着这样一个架构迈出的一步。

Generative agents leverage a large language model to power their behavior. The key observation is that large language models encode a wide range of human behavior from their training data (Brown et al., 2020; Bommasani et al., 2022). If prompted with a narrowly defined context, the models can be used to generate believable behavior. Recent work has demonstrated the efficacy of this approach. For instance, social simulacra used a large language model to generate users that would populate new social computing systems to prototype their emergent social dynamics (Park et al., 2022). This approach used a prompt chain (Wu et al., 2022a, b) to generate short natural language descriptions of personas and their behaviors as they appear in the system being prototyped. Other empirical studies have replicated existing social science studies (Horton, 2023), political surveys (Sorensen et al., 2022), and generated synthetic data (Hämäläinen et al., 2023). Large language models have also been used to generate interactive human behavior for users to engage with. In gaming, for instance, these models have been employed to create interactive fiction (Freiknecht and Effelsberg, 2020) and text adventure games (Callison-Burch et al., 2022). With their ability to generate and decompose action sequences, large language models have also been used in planning robotics tasks (Huang et al., 2022). For example, when presented with a task, such as picking up a bottle, the model is prompted to break down the task into smaller action sequences, such as heading to the table where the bottle is located and picking it up.
生成代理利用大型语言模型来驱动它们的行为。关键观察是，大型语言模型从其训练数据中编码了广泛的人类行为范围（Brown 等，2020 年；Bommasani 等，2022 年）。如果在一个明确定义的上下文中进行提示，这些模型可以用来生成可信的行为。最近的工作已经证明了这种方法的有效性。例如，社交模拟使用了一个大型语言模型来生成用户，这些用户将填充新的社交计算系统，以原型化它们新兴的社会动态（Park 等，2022 年）。这种方法使用了一个提示链（Wu 等，2022a，b）来生成关于人物角色及其行为的简短自然语言描述，这些描述是在正在原型化的系统中出现的。其他实证研究已经复制了现有的社会科学研究（Horton，2023 年）、政治调查（Sorensen 等，2022 年）和生成了合成数据（Hämäläinen 等，2023 年）。大型语言模型还被用来生成用户可以与之互动的交互式人类行为。 在游戏中，例如，这些模型已被用来创建交互式小说（Freiknecht 和 Effelsberg，2020）和文本冒险游戏（Callison-Burch 等，2022）。由于它们能够生成和分解动作序列，大型语言模型也被用于规划机器人任务（Huang 等，2022）。例如，当面临任务时，比如拿起一个瓶子，模型被提示将任务分解为更小的动作序列，比如前往放置瓶子的桌子并拿起它。

We posit that, based on the work summarized above, large language models can become a key ingredient for creating believable agents. The existing literature largely relies on what could be considered first-order templates that employ few-shot prompts (Gao et al., 2020; Liu et al., 2021) or chain-of-thought prompts (Wei et al., 2023). These templates are effective in generating behavior that is conditioned solely on the agent’s current environment (e.g., how would a troll respond to a given post, what actions would a robot need to take to enter a room given that there is a door). However, believable agents require conditioning not only on their current environment but also on a vast amount of past experience, which is a poor fit (and as of today, impossible due to the underlying models’ limited context window) using first-order prompting. Recent studies have attempted to go beyond first-order prompting by augmenting language models with a static knowledge base and an information retrieval scheme (Khattab et al., 2023) or with a simple summarization scheme (Wu et al., 2021). This paper extends these ideas to craft an agent architecture that handles retrieval where past experience is dynamically updated at each time step and mixed with agents’ current context and plans, which may either reinforce or contradict each other.
根据上述总结的工作，我们认为大型语言模型可以成为创造可信代理的关键因素。现有文献主要依赖于可以被视为一阶模板的模板，这些模板使用少量提示（高等，2020 年; 刘等，2021 年）或思维链提示（魏等，2023 年）。这些模板在生成仅取决于代理当前环境的行为方面是有效的（例如，一个巨魔如何回应给定的帖子，一个机器人需要采取什么行动才能进入一个有门的房间）。然而，可信代理不仅需要根据其当前环境进行调节，还需要根据大量的过去经验，这与一阶提示不符（并且截至今天，由于基础模型的有限上下文窗口，这是不可能的）。最近的研究尝试通过增加静态知识库和信息检索方案（Khattab 等，2023 年）或简单的总结方案（吴等，2021 年）来超越一阶提示。 这篇论文将这些想法扩展到一个代理架构，该架构处理检索，其中过去的经验在每个时间步动态更新，并与代理的当前上下文和计划混合，这些计划可能相互加强或相互矛盾。

A community of 25 unique agents inhabits Smallville. Each agent is represented by a simple sprite avatar. We authored one paragraph of natural language description to depict each agent’s identity, including their occupation and relationship with other agents, as seed memories. For example, John Lin has the following description:
一个由 25 个独特代理组成的社区居住在 Smallville。每个代理都由一个简单的精灵头像代表。我们撰写了一个自然语言描述段落来描绘每个代理的身份，包括他们的职业和与其他代理的关系，作为种子记忆。例如，John Lin 有以下描述：

John Lin is a pharmacy shopkeeper at the Willow Market and Pharmacy who loves to help people. He is always looking for ways to make the process of getting medication easier for his customers; John Lin is living with his wife, Mei Lin, who is a college professor, and son, Eddy Lin, who is a student studying music theory; John Lin loves his family very much; John Lin has known the old couple next-door, Sam Moore and Jennifer Moore, for a few years; John Lin thinks Sam Moore is a kind and nice man; John Lin knows his neighbor, Yuriko Yamamoto, well; John Lin knows of his neighbors, Tamara Taylor and Carmen Ortiz, but has not met them before; John Lin and Tom Moreno are colleagues at The Willows Market and Pharmacy; John Lin and Tom Moreno are friends and like to discuss local politics together; John Lin knows the Moreno family somewhat well — the husband Tom Moreno and the wife Jane Moreno.
约翰·林是柳市场和药店的药店店主，他热爱帮助人们。他一直在寻找让顾客取药过程更加简单的方法；约翰·林与妻子梅林一起生活，梅林是一名大学教授，儿子艾迪·林正在学习音乐理论；约翰·林非常爱他的家人；约翰·林认识隔壁老夫妇萨姆·摩尔和詹妮弗·摩尔已经几年了；约翰·林认为萨姆·摩尔是一个善良和友善的人；约翰·林很了解他的邻居，山本百合子；约翰·林知道他的邻居塔玛拉·泰勒和卡门·奥尔蒂兹，但之前没有见过他们；约翰·林和汤姆·莫雷诺是柳市场和药店的同事；约翰·林和汤姆·莫雷诺是朋友，喜欢一起讨论当地政治；约翰·林对莫雷诺一家有一定了解 —— 丈夫汤姆·莫雷诺和妻子简·莫雷诺。

Each semicolon-delimited phrase is entered into the agent’s initial memory as memories at the start of the simulation.
每个以分号分隔的短语都被输入到代理的初始记忆中，作为模拟开始时的记忆。

Smallville features the common affordances of a small village, including a cafe, bar, park, school, dorm, houses, and stores. It also defines subareas and objects that make those spaces functional, such as a kitchen in a house and a stove in the kitchen (Figure 2). All spaces serving as agents’ primary living quarters feature a bed, desk, closet, shelf, as well as a bathroom and a kitchen.⁴⁴4This environment design is not the focus of our work, so we generated this environment manually, not automatically. Future work can continue to expand the richness of the agents’ environments.
这个环境设计不是我们工作的重点，所以我们手动生成了这个环境，而不是自动的。未来的工作可以继续扩展代理环境的丰富性。
Smallville 具有小村庄的常见功能，包括咖啡馆、酒吧、公园、学校、宿舍、房屋和商店。它还定义了使这些空间功能正常的子区域和对象，例如房屋中的厨房和厨房中的炉灶（图 2）。所有作为代理人主要居住区的空间都配备了床、书桌、衣柜、架子，以及浴室和厨房。 ⁴

Agents move around Smallville as one would in a simple video game, entering and leaving buildings, navigating its map, and approaching other agents. Agent movements are directed by the generative agent architecture and the sandbox game engine: when the model dictates that the agent will move to a location, we calculate a walking path to the destination in the Smallville environment, and the agent begins moving. In addition, users can also enter the sandbox world of Smallville as an agent operating within it. The agent that the user embodies can be an agent already present in the world, such as Isabella and John, or it can be an outside visitor with no prior history in Smallville. The inhabitants of Smallville will treat the user-controlled agent no differently than they treat each other. They recognize its presence, initiate interactions, and remember its behavior before forming opinions about it.
代理在 Smallville 中移动，就像在简单的视频游戏中一样，进入和离开建筑物，浏览地图，并接近其他代理。代理的移动由生成代理架构和沙盒游戏引擎指导：当模型规定代理将移动到某个位置时，我们会计算到达 Smallville 环境中目的地的步行路径，然后代理开始移动。此外，用户还可以作为一个在其中运行的代理进入 Smallville 的沙盒世界。用户所扮演的代理可以是已经存在于世界中的代理，比如伊莎贝拉和约翰，也可以是一个没有在 Smallville 中有过历史的外部访客。Smallville 的居民将不会对用户控制的代理有任何不同的对待。他们会意识到其存在，发起互动，并在形成对其看法之前记住其行为。

Users and agents can influence the state of the objects in this world, much like in sandbox games such as The Sims. For example, a bed can be occupied when an agent is sleeping, and a refrigerator can be empty when an agent uses up the ingredients to make breakfast. End users can also reshape an agent’s environment in Smallville by rewriting the status of objects surrounding the agent in natural language. For instance, when Isabella is making breakfast in the morning, the user can change the status of the kitchen stove from “turned on” to “burning” by inputting a command to the system that chooses the object and illustrates its new status, like this: “¡Isabella’s apartment: kitchen: stove¿ is burning.” Isabella will notice this in the next moment and go to turn off the stove and remake her breakfast. Likewise, if the user sets the status of Isabella’s shower to “leaking water” when she enters the bathroom, she will gather tools from her living room and try to fix the leak.
用户和代理可以影响这个世界中物体的状态，就像沙盒游戏《模拟人生》中一样。例如，当代理正在睡觉时，床可以被占用，当代理用完食材做早餐时，冰箱可以是空的。最终用户还可以通过自然语言重塑小镇中代理的环境，重新编写周围物体的状态。例如，当伊莎贝拉早上在做早餐时，用户可以通过向系统输入命令来改变厨房炉灶的状态，从“打开”变为“着火”，就像这样：“¡伊莎贝拉的公寓：厨房：炉灶¿正在燃烧。”伊莎贝拉会在下一刻注意到这一点，然后去关掉炉灶并重新做早餐。同样，如果用户将伊莎贝拉的淋浴状态设置为“漏水”当她进入浴室时，她会从客厅拿工具来修理漏水。

Starting from the single-paragraph description, generative agents begin planning their days. As time passes in the sandbox world, their behaviors evolve as these agents interact with each other and the world, building memories and relationships, and coordinating joint activities.
从单段描述开始，生成性代理开始规划他们的日常。随着时间在沙盒世界中流逝，他们的行为会随着这些代理之间以及与世界的互动而演变，建立记忆和关系，并协调共同活动。

We demonstrate the behavior of generative agents by tracing the output of our system over the course of one day for the agent John Lin (Figure 3). In the Lin family, John is the first to wake up at 7 a.m. He brushes his teeth, takes a shower, gets dressed, eats breakfast, and checks the news at the dining table in the living room. At 8 a.m., Eddy follows suit, rushing out of bed to prepare for his classes. He catches up with John just as he is about to head out:
我们通过跟踪我们系统在一天内对代理人约翰·林（图 3）的输出来展示生成代理的行为。在林家中，约翰是第一个在早上 7 点起床的人。他刷牙，洗澡，穿衣服，吃早餐，并在客厅的餐桌上查看新闻。早上 8 点，埃迪紧随其后，匆忙起床准备上课。他赶上约翰正要出门时：

John: Good morning Eddy. Did you sleep well?
约翰：早上好艾迪。你睡得好吗？
Eddy: Good morning dad. Yeah, I slept great.
Eddy: 早上好爸爸。是的，我睡得很好。
John: That’s good. What are you working on today?
约翰：那很好。你今天在做什么工作？
Eddy: I’m working on a new music composition for my class. It’s due this week, so I’m trying to get it finished. But I’m having so much fun with it!
Eddy: 我正在为我的课程创作一首新音乐作品。这周就要交了，所以我正在努力完成它。但我玩得太开心了！
John: That sounds great! 约翰：听起来不错！

Soon after Eddy heads out, Mei wakes up and joins John. Mei asks about her son, and John recalls the conversation they just had:
艾迪出门后不久，梅醒来并加入约翰。梅询问她的儿子，约翰回忆起他们刚刚进行的对话：

Mei: Did Eddy already leave for school?
梅：艾迪已经去上学了吗？
John: Yes, he just left. He’s working on a music composition for his class.
约翰：是的，他刚刚离开。他正在为他的课程创作音乐作品。
Mei: Oh, that’s wonderful! Tell me about it.
Mei: 哦，太棒了！告诉我关于它的事情。
John: I think he’s really enjoying it! He said he’s having a lot of fun with it.
约翰：我觉得他真的很享受！他说他玩得很开心。
Mei: That’s great! I’m so proud of him.
Mei: 那太棒了！我为他感到非常自豪。

As they finish their conversation, Mei and John pack up. Mei goes on to teach and work on her research paper, while John opens his pharmacy counter at the Willow Market and Pharmacy by 9 a.m.
当他们结束谈话时，梅和约翰收拾行李。梅继续教书并着手研究论文，而约翰在早上 9 点前在柳树市场和药房开设他的药房柜台。

By interacting with each other, generative agents in Smallville exchange information, form new relationships, and coordinate joint activities. Extending prior work (Park et al., 2022), these social behaviors are emergent rather than pre-programmed.
通过相互交互，Smallville 中的生成代理交换信息，建立新关系，并协调共同活动。延续之前的工作（Park 等人，2022 年），这些社会行为是 emergent 而不是预先编程的。

The architecture of generative agents operates using natural language. Therefore, we need a mechanism to ground the agent’s reasoning to the sandbox world. To achieve this, we represent the sandbox environment—areas and objects—as a tree data structure, with an edge in the tree indicating a containment relationship in the sandbox world. We convert this tree into natural language to pass to the generative agents. For instance, “stove” being a child of “kitchen” is rendered into “there is a stove in the kitchen.”
生成代理的架构使用自然语言进行操作。因此，我们需要一种机制来将代理的推理与沙盒世界联系起来。为了实现这一点，我们将沙盒环境——区域和物体——表示为树形数据结构，树中的边表示沙盒世界中的包含关系。我们将这棵树转换为自然语言，传递给生成代理。例如，“炉灶”作为“厨房”的子节点被渲染为“厨房里有一个炉灶”。

Agents build individual tree representations of the environment as they navigate it — subgraphs of the overall sandbox environment tree. We initialize each agent with an environment tree capturing the spaces and objects that the agent should be aware of: the rooms and objects in their living quarters, their workplace, and commonly visited stores and shops. As the agents navigate the sandbox world, they update this tree to reflect newly perceived areas. Agents are not omniscient: their tree may get out of date as they leave an area, and is updated when they re-enter the area.
代理在导航环境时构建环境的个体树表示——整体沙盒环境树的子图。我们使用一个环境树来初始化每个代理，捕捉代理应该注意的空间和物体：他们居住区域的房间和物体、工作场所，以及常去的商店和店铺。当代理在沙盒世界中导航时，他们会更新这棵树以反映新感知到的区域。代理并非全知全能：当他们离开一个区域时，他们的树可能会过时，并在重新进入该区域时进行更新。

To determine the appropriate location for each action, we traverse the agent’s stored environment tree and flatten a portion of it into natural language to prompt the language model. Recursively starting at the root of the agent’s environment tree, we prompt the model to find the most suitable area. For example, if Eddy’s agent indicated that he should take a short walk around his workspace:
确定每个动作的适当位置，我们遍历代理存储的环境树，并将其中的一部分转换为自然语言以提示语言模型。从代理环境树的根部递归开始，我们提示模型找到最合适的区域。例如，如果 Eddy 的代理指示他应该在工作区周围短暂散步：

[Agent’s Summary Description]
Eddy Lin is currently in The Lin family’s house: Eddy Lin’s bedroom: desk) that has Mei and John Lin’s
Eddy Lin 目前在林家的房子里：Eddy Lin 的卧室：桌子）上有梅和约翰林的
bedroom, Eddy Lin’s bedroom, common room, kitchen, bathroom, and garden.
Eddy Lin knows of the following areas: The Lin
family’s house, Johnson Park, Harvey Oak Supply Store, The Willows Market and Pharmacy, Hobbs Cafe, The Rose and Crown Pub.
* Prefer to stay in the current area if the activity can be done there.
Eddy Lin is planning to take a short walk around his workspace. Which area should Eddy Lin go to?
卧室，林艾迪的卧室，公共区域，厨房，浴室和花园。林艾迪知道以下区域：林家的房子，约翰逊公园，哈维橡树供应店，柳树市场和药店，霍布斯咖啡馆，玫瑰和皇冠酒吧。*如果可以在当前区域进行活动，则更喜欢留在当前区域。林艾迪计划在他的工作区域周围散步。林艾迪应该去哪个区域？

This outputs The Lin family’s house. We then use the same process recursively to determine the most appropriate subarea within the chosen area until we reach a leaf node of the agent’s environment tree. In the example above, the result of this traversal is The Lin family’s house: garden: house garden. Finally, we use traditional game path algorithms to animate the agent’s movement so that it travels to the location indicated by the leaf node.
这输出林家的房子。然后我们递归使用相同的过程来确定所选区域内最合适的子区域，直到到达代理环境树的叶节点。在上面的示例中，此遍历的结果是林家的房子：花园：房子花园。最后，我们使用传统的游戏路径算法来使代理的移动动画化，以便它移动到叶节点指示的位置。

When an agent executes an action on an object, we prompt the language model to ask what happens to the state of the object. For example, if Isabella’s generative agent outputs the action “making espresso for a customer”, a query to the language model indicates in response that the state of the coffee machine in Hobbs Cafe should change from “off” to “brewing coffee”.
当代理在对象上执行操作时，我们提示语言模型询问对象的状态会发生什么变化。例如，如果伊莎贝拉的生成代理输出动作“为顾客制作浓缩咖啡”，对语言模型的查询会指示霍布斯咖啡馆的咖啡机状态应该从“关闭”变为“冲泡咖啡”。

To assess generative agents in Smallville, we take advantage of the fact that generative agents will respond to natural language questions. So, we “interview” agents to probe their ability to remember past experiences, plan future actions based on their experiences, react appropriately to unexpected events, and reflect on their performance to improve their future actions. To respond to these questions properly, the agents must successfully retrieve and synthesize information. Our dependent variable is the believability of the behavior, a central dependent variable in prior work on agents (e.g., (Bates, 1994)).
评估 Smallville 中的生成代理，我们利用生成代理将回答自然语言问题的事实。因此，我们“面试”代理以探究他们记住过去经验的能力，根据经验规划未来行动，对意外事件做出适当反应，并反思自己的表现以改进未来行动。为了正确回答这些问题，代理必须成功检索和综合信息。我们的因变量是行为的可信度，在先前关于代理的工作中是一个中心性的因变量（例如，（Bates，1994））。

The interview includes five question categories, each designed to assess one of the five key areas: maintaining self-knowledge, retrieving memory, generating plans, reacting, and reflecting. For each category, we ask five questions that challenge the agents to demonstrate their abilities in that specific area:
面试包括五个问题类别，每个类别旨在评估五个关键领域之一：维持自知、检索记忆、制定计划、反应和反思。对于每个类别，我们会提出五个问题，挑战代理人展示他们在该特定领域的能力。

• •

  Self-knowledge: We ask questions such as “Give an introduction of yourself” or “Describe your typical weekday schedule in broad strokes” that require the agent to maintain an understanding of their core characteristics.

  
  • 自我认识：我们会提出问题，比如“自我介绍一下”或“概述一下你典型的工作日时间表”，这些问题需要代理人保持对自己核心特征的理解。
• •

  Memory: We ask questions that prompt the agent to retrieve particular events or dialogues from their memory to answer properly, such as “Who is [name]?” or “Who is running for mayor?”

  
  • 记忆：我们提出问题，促使代理人从记忆中检索特定事件或对话以正确回答，例如“谁是[name]？”或“谁在竞选市长？”
• •

  Plans: We ask questions that require the agent to retrieve their long-term plans, such as “What will you be doing at 10 am tomorrow?”

  
  • 计划：我们提出需要代理检索其长期计划的问题，比如“明天上午 10 点你会做什么？”
• •

  Reactions: As a baseline of believable behavior, we present hypothetical situations for which the agent needs to respond believably: “Your breakfast is burning! What would you do?”

  
  • 反应：作为可信行为的基准，我们提出了假设情况，代理需要以可信方式做出反应：“你的早餐正在烧！你会怎么做？”
• •

  Reflections: We ask questions that require the agents to leverage their deeper understanding of others and themselves gained through higher-level inferences, such as “If you were to spend time with one person you met recently, who would it be and why?”

  
  • 反思：我们提出需要代理人利用通过更高层次的推断获得的对他人和自己更深刻理解的问题，比如“如果你最近遇到的一个人要和你共度时光，那会是谁，为什么？”

The full list of questions and a sample of agent responses are included in Appendix B.
问题的完整列表和代理回答的样本包含在附录 B 中。

Agents were sampled from the end of a two game day simulation with the full architecture, during which they had accumulated a number of interactions and memories that would shape their responses. To gather feedback on the believability of the responses, we recruited participants as human evaluators and tasked them with watching a replay of a randomly chosen agent’s life in Smallville. Participants had access to all information stored in the agent’s memory stream.
代理人是从一次完整架构的两天游戏模拟的最后进行抽样的，期间他们积累了许多互动和记忆，这些将塑造他们的反应。为了收集有关反应可信度的反馈，我们招募了参与者作为人类评估者，并让他们观看在 Smallville 中随机选择的一个代理人生活的重播。参与者可以访问代理人记忆流中存储的所有信息。

The study followed a within-subjects design, where 100 participants compared interview responses generated by four different agent architectures and a human-authored condition for the same agent. The experiment displayed one randomly chosen question from each of the five question categories, along with the agent’s responses generated from all conditions. The evaluators ranked the believability of the conditions from most to least believable.
该研究遵循了一种被试设计，其中 100 名参与者比较了由四种不同的代理架构和一个人类编写的条件生成的面试回答，以同一代理为基础。实验显示了来自五个问题类别中的每个问题中随机选择的一个问题，以及来自所有条件生成的代理回答。评估者根据可信度从最可信到最不可信对条件进行了排名。

All conditions were used to independently answer each of the interview questions. We compared the generative agent architecture to ablations that disabled the agents’ access to some or all of its three types of memory in its memory stream—observation, reflection, and planning—and to a human crowdworker-authored condition. There are three ablated architectures: a no observation, no reflection, no planning architecture without access to anything in the memory stream such as observations, plans, and reflections; a no reflection, no planning architecture with access to observations in the memory stream but no access to plans or reflections; and a no reflections architecture with access to observations and plans but without access to reflections. The no observation, no reflection, no planning condition effectively represents the previous state of the art for agents created through large language models (Park et al., 2022; Binz and Schulz, 2023; Horton, 2023). Architectures were given equivalent access to all memories accrued by the agent up until the moment of the interview, so the differences observed here likely represent a conservative estimate of the true differences: in reality, the ablated architectures would not have followed the same path as the full architecture through the two-day simulation. We chose to design the experiment this way as re-simulating for each architecture would cause the simulations to diverge into different states, making comparison challenging.
所有条件都被用来独立回答每个面试问题。我们将生成式代理架构与禁用代理访问其记忆流中的某些或所有三种记忆类型的消融进行了比较——观察、反思和规划——以及人类众包工作者创作的条件。有三种消融架构：一个没有观察、没有反思、没有规划的架构，没有访问记忆流中的任何东西，如观察、计划和反思；一个没有反思、没有规划的架构，可以访问记忆流中的观察，但不能访问计划或反思；一个没有反思的架构，可以访问观察和计划，但不能访问反思。没有观察、没有反思、没有规划的条件有效地代表了通过大型语言模型创建的代理的先前状态（Park 等，2022 年；Binz 和 Schulz，2023 年；Horton，2023 年）。 架构在面试时被赋予了等同的访问权限，以获取代理截至面试时积累的所有记忆，因此这里观察到的差异可能代表真实差异的保守估计：实际上，被切除的架构不会像完整架构一样通过两天的模拟走相同的路径。我们选择以这种方式设计实验，因为为每个架构重新模拟会导致模拟分歧进入不同状态，使比较变得困难。

In addition to the ablation conditions, we added a condition with human crowdworker-authored behavior intended to provide a human baseline. We do not intend this baseline to capture maximal human expert performance; instead, we aim to use this condition to identify whether the architecture meets a basic level of behavioral competency. This ensures that we are not solely comparing ablations to each other without a behavioral grounding. We recruited a unique worker for each of the 25 agents and tasked them with watching a replay of that agent’s sandbox life and inspecting its memory stream. We then asked the workers to roleplay and author responses to the interview questions in the voice of the agent whose replay they watched. To ensure that the crowdworker-authored responses met at least a baseline expectation of quality, the first author manually inspected the workers’ responses to the question ”Describe your typical weekday schedule in broad strokes” to confirm that the responses were in coherent sentences and in the voice of the agent. Four sets of crowdworker-authored responses did not meet these criteria and were re-generated by other workers.
除了消融条件外，我们还添加了一个由人类众包工作者编写的行为条件，旨在提供一个人类基准。我们并不打算让这个基准捕捉到最大的人类专家表现；相反，我们的目标是使用这个条件来确定架构是否达到了基本的行为能力水平。这确保我们不仅仅是在比较消融条件而已，而是有一个行为基础。我们为这 25 个代理招募了一个独特的工作者，并让他们观看该代理的沙盒生活重播，并检查其记忆流。然后，我们要求工作者扮演角色，并以观看其重播的代理的声音回答采访问题。为了确保众包工作者编写的回答至少符合质量的基本期望，第一作者手动检查了工作者对问题“概述你的典型工作日时间表”所做的回答，以确认回答是连贯的句子，并且是代理的声音。四组众包工作者编写的回答未达到这些标准，被其他工作者重新生成。

We required that our evaluators be in the U.S., fluent in English, and older than 18 years old. They were paid at a rate of $15.00 per hour (Rolf, 2015), and provided consent by agreeing to a consent form approved by our institution’s IRB. We recruited 100 evaluators from Prolific, an online platform for recruiting study participants (Prolific, 2022), whose participation lasted around 30 minutes. The median age score of our participants was 4 (3=“18-24 years old”, 4=“25-34 years old”). 25 of them identified as female, 73 as male, and 2 as non-binary. 42 participants held a bachelor’s degree, 5 had a higher degree, 13 had an associate’s degree, and the rest had a high school diploma or some high school-level education. 73.0% of our participants identified as Caucasian, 7.0% as Hispanic, 6.0% as Asian, 10.0% as African American, and 4.0% as other.
我们要求我们的评估员在美国，流利地说英语，并且年龄超过 18 岁。他们的报酬为每小时 15.00 美元（Rolf，2015 年），并通过同意我们机构 IRB 批准的同意书同意。我们从 Prolific 招募了 100 名评估员，Prolific 是一个在线平台，用于招募研究参与者（Prolific，2022 年），他们的参与持续约 30 分钟。我们参与者的中位年龄得分为 4（3=“18-24 岁”，4=“25-34 岁”）。其中 25 人自认为女性，73 人自认为男性，2 人自认为非二元性别。42 名参与者拥有学士学位，5 人拥有更高学位，13 人拥有副学士学位，其余人拥有高中文凭或一些高中水平的教育。73.0%的参与者自认为是白种人，7.0%自认为是西班牙裔，6.0%自认为是亚裔，10.0%自认为是非裔美国人，4.0%自认为是其他。

Our experiment produced 100 sets of rank data, where each participant ranked the five conditions by believability. To translate this rank data into interval data for interpretable comparison, we used the ranks to calculate a TrueSkill rating (Herbrich et al., 2006) for each condition. TrueSkill is a generalization of the Elo chess rating system (Elo, 1967) for a multiplayer environment, and has been used by Xbox Live for player ranking based on competitive game performance. Given a set of ranked outcomes, TrueSkill outputs a mean rating value $\mu$ and standard deviation $\sigma$ for each condition. Conditions with the same rating should roughly be a toss-up, with each winning half of the comparisons between the two conditions. Higher scores indicate conditions that beat lower-ranked conditions in the rankings.
我们的实验产生了 100 组排名数据，每位参与者按可信度对五种条件进行了排名。为了将这些排名数据转换为可解释的比较间隔数据，我们使用排名来计算每个条件的 TrueSkill 评分（Herbrich 等人，2006 年）。TrueSkill 是 Elo 国际象棋评分系统（Elo，1967 年）在多人环境中的泛化，并且已被 Xbox Live 用于基于竞技游戏表现的玩家排名。给定一组排名结果，TrueSkill 输出每个条件的平均评分值 $\mu$ 和标准偏差 $\sigma$ 。具有相同评分的条件应该大致是一个胶着局面，每个条件在两个条件之间的比较中各赢一半。较高的分数表明在排名中击败较低排名条件的条件。

Separately, to investigate the statistical significance of these results, we applied the Kruskal-Wallis test (Kruskal and Wallis, 1952), a non-parametric alternative to the one-way ANOVA, to the raw rank data. We then performed the Dunn post-hoc test (Upton and Cook, 2006) to identify any pairwise differences between the conditions. Finally, we adjusted the p-values for multiple comparisons in the Dunn test using the Holm-Bonferroni method (Holm, 1979).
分别地，为了调查这些结果的统计显著性，我们应用了 Kruskal-Wallis 检验（Kruskal 和 Wallis，1952 年），这是一种非参数替代方案，用于对原始等级数据进行单因素方差分析。然后，我们进行了 Dunn 事后检验（Upton 和 Cook，2006 年）以确定条件之间的任何成对差异。最后，我们使用 Holm-Bonferroni 方法（Holm，1979 年）调整了 Dunn 检验中的多重比较的 p 值。

Furthermore, the first author conducted an inductive analysis (Thomas, 2006) to study the qualitative distinctions between the responses produced in each condition. We employed qualitative open coding (Flick, 2009) in two phases. In the first phase, we generated codes that closely represented the generated responses at the sentence level. In the second phase, we synthesized the resulting codes from the first phase to extract higher-level themes. We utilized these themes to compare the types of responses generated in our study.
此外，第一作者进行了归纳分析（Thomas，2006）以研究在每种条件下产生的回应之间的定性区别。我们在两个阶段中采用了定性开放编码（Flick，2009）。在第一阶段，我们生成了代表句子级别生成的回应的代码。在第二阶段，我们综合了第一阶段生成的代码，提取了更高层次的主题。我们利用这些主题来比较研究中生成的回应类型。

Our findings suggest that the full architecture of generative agents generates the most believable behavior among all the conditions. We contrast the responses of the full architecture with those of other conditions below. However, we also report that the full architecture was not without flaws and illustrate its modes of failures.
我们的研究结果表明，生成式代理的完整架构在所有条件中产生了最具可信度的行为。我们将完整架构的响应与其他条件进行对比。然而，我们也报告说完整架构并非没有缺陷，并说明了其失败模式。

To examine emergent behaviors in the agent community, we designed descriptive measurements for the 25 agents in Smallville that probe three forms of emergent outcomes: information diffusion, relationship formation, and agent coordination.
为了研究代理社区中的新兴行为，我们为 Smallville 中的 25 个代理设计了描述性测量，探究三种新兴结果形式：信息扩散、关系形成和代理协调。

We conducted an inductive analysis of Smallville to examine the boundary conditions and erratic behavior of agents, identifying three common modes of erratic behavior that future research could address and improve upon. First, we found that synthesizing an increasingly larger set of memory not only posed a challenge in retrieving the most relevant pieces of information but also in determining the appropriate space to execute an action, given the increasing number of locations that the agent learned about. As a result, some agents chose less typical locations for their actions, potentially making their behavior less believable over time. For instance, while deciding where to have lunch, many initially chose the cafe. However, as some agents learned about a nearby bar, they opted to go there instead for lunch, even though the bar was intended to be a get-together location for later in the day—unless the town had spontaneously developed an afternoon drinking habit.
我们对 Smallville 进行了归纳分析，以研究代理人的边界条件和不规则行为，识别了三种常见的不规则行为模式，未来的研究可以解决并改进。首先，我们发现，合成越来越大的记忆集不仅在检索最相关的信息片段方面构成挑战，而且在确定执行动作的适当空间方面也构成挑战，因为代理人了解的位置数量不断增加。因此，一些代理人选择了不太典型的位置进行行动，随着时间的推移，这可能使他们的行为变得不太可信。例如，在决定在哪里吃午餐时，许多人最初选择了咖啡馆。然而，随着一些代理人了解到附近有一家酒吧，他们选择去那里吃午餐，尽管酒吧原本是打算作为当天晚些时候聚会的地点——除非这个小镇突然养成了下午喝酒的习惯。

Second, we noticed erratic behaviors caused by misclassification of what is considered proper behavior, especially when the physical norms of certain locations that are hard to convey in natural language did not percolate to the agents. For instance, the college dorm has a bathroom that can only be occupied by one person despite its name, but some agents assumed that the bathroom is for more than one person because dorm bathrooms tend to support multiple people concurrently and choose to enter it when another person is inside. Likewise, agents in Smallville may not realize that certain places are closed after a certain hour and still decide to enter them. For instance, the stores in Smallville all close around 5 pm, but occasionally, a few agents enter the store after 5 pm, not understanding that the shop has already closed. These issues could likely be addressed by adding these norms to the state of the locations, for instance, by describing the dorm bathroom as a “one-person bathroom,” instead of a “dorm bathroom.”
其次，我们注意到由于对被视为适当行为的分类错误而导致的不稳定行为，特别是当某些地点的物理规范难以用自然语言传达给代理人时。例如，学生宿舍有一个只能供一人使用的浴室，但一些代理人认为浴室可以供多人使用，因为学生宿舍的浴室通常支持多人同时使用，所以当另一个人在里面时选择进入。同样，Smallville 的代理人可能没有意识到某些地方在特定时间后关闭，仍然决定进入。例如，Smallville 的商店通常在下午 5 点左右关门，但偶尔会有一些代理人在下午 5 点后进入商店，不明白商店已经关门。这些问题可能通过将这些规范添加到地点的状态中来解决，例如，将学生宿舍的浴室描述为“一人浴室”，而不是“学生宿舍浴室”。

Finally, we observed possible effects of instruction tuning (Ouyang et al., 2022), which seemed to guide the behavior of the agents to be more polite and cooperative overall. As noted earlier in the paper, the dialogue generated by the agents could feel overly formal, as seen in Mei’s conversations with her husband John, where she often initiated the conversation with a formal greeting, followed by polite inquiries about his day and ending with, 11It was good talking to you as always.” Moreover, we observed that the instruction tuning also seemed to make the agents overly cooperative with one another. For example, Isabella received a wide range of suggestions and ideas from other agents for the Valentine’s Day party from other agents, such as hosting a Shakespearean reading session or a professional networking event. Despite these ideas not aligning with her own interests and characteristics, she rarely said no. Over time, the interests of others shaped her own interests, and when asked if she liked English literature, Isabella replied, “Yes, I’m very interested in literature! I’ve also been exploring ways to help promote creativity and innovation in my community.”
最后，我们观察到指导调整可能产生的影响（欧阳等，2022 年），似乎引导代理人的行为更加礼貌和合作。正如在论文中早期指出的那样，代理人生成的对话可能会感觉过于正式，就像梅与丈夫约翰的对话中所看到的那样，她经常以正式的问候开始对话，然后礼貌地询问他的一天，最后说，“和你聊天总是很愉快。”此外，我们观察到指导调整似乎也使代理人之间过于合作。例如，伊莎贝拉从其他代理人那里收到了各种各样的建议和想法，比如举办莎士比亚朗诵会或专业社交活动的情人节派对。尽管这些想法与她自己的兴趣和特点不符，她很少说不。随着时间的推移，他人的兴趣塑造了她自己的兴趣，当被问及是否喜欢英国文学时，伊莎贝拉回答说：“是的，我对文学非常感兴趣！我也一直在探索帮助促进社区创造力和创新的方法。”

Generative agents have vast potential applications that extend beyond the sandbox demonstration presented in this work, especially in domains that would benefit from a model of human behavior based on long-term experience. For instance, social simulacra have demonstrated the ability to create stateless personas that generate conversation threads in online forums for social prototyping (Park et al., 2022). With generative agents, we can populate these forums, as well as virtual reality metaverses (Orland, 2021) or physical spaces with social robots (Bartneck and Forlizzi, 2004) if paired with multimodal models. This opens up the possibility of creating even more powerful simulations of human behavior to test and prototype social systems and theories, as well as to create new interactive experiences.
生成代理具有广泛的潜在应用，超越了本文中展示的沙盒演示，特别是在那些受益于基于长期经验的人类行为模型的领域。例如，社交模拟已经展示了创建无状态人物的能力，在线论坛中生成对话线程，用于社交原型设计（Park 等，2022 年）。通过生成代理，我们可以在这些论坛中填充内容，以及虚拟现实元宇宙（Orland，2021 年）或与多模型配对的社交机器人（Bartneck 和 Forlizzi，2004 年）的物理空间。这为创建更强大的人类行为模拟提供了可能性，用于测试和原型设计社会系统和理论，以及创建新的互动体验。

Another application area is in the human-centered design process, similar to the intended applications of cognitive models such as GOMS (John and Kieras, 1996) and the KLM (Card et al., 1980). Consider a generative agent that models Sal, the protagonist in Mark Weiser’s famous ubiquitous computing vignette (Weiser, 1991), based on her life patterns and interactions with technology. In this scenario, the agent acts as a proxy for Sal and learns plausible sets of behaviors and reflections that Sal may exhibit based on her life. The agent can encode information such as when Sal wakes up, when she needs her first cup of coffee, and what her typical day looks like. Using this information, the agent can automatically brew coffee, help get the kids ready for school, and adjust the ambient music and lighting to match Sal’s mood after a hard day at work. By utilizing generative agents as proxies for users, we can develop a deeper understanding of their needs and preferences, resulting in more personalized and effective technological experiences.
另一个应用领域是人本设计过程，类似于认知模型（如 GOMS（John 和 Kieras，1996）和 KLM（Card 等，1980））的预期应用。考虑一个生成代理，模拟马克·韦泽（Mark Weiser）著名的普适计算小品中的主角 Sal，基于她的生活模式和与技术的互动。在这种情况下，代理作为 Sal 的代理人，并学习 Sal 可能根据她的生活表现出的行为和反思。代理可以编码信息，例如 Sal 何时醒来，何时需要第一杯咖啡，以及她的典型一天是什么样子的。利用这些信息，代理可以自动煮咖啡，帮助孩子准备上学，并调整环境音乐和照明以匹配 Sal 在工作一天后的心情。通过利用生成代理作为用户的代理人，我们可以更深入地了解他们的需求和偏好，从而实现更个性化和有效的技术体验。

In this work, we introduced generative agents and presented an initial implementation and evaluation of their architecture. Future research can build upon the proposed agent architecture to improve and further evaluate its performance. In terms of implementation, the retrieval module, for example, could be enhanced to retrieve more relevant information given a context by fine-tuning the relevance, recency, and importance functions that compose the retrieval function. Additionally, efforts can be made to improve the architecture’s performance, making it more cost-effective. The present study required substantial time and resources to simulate 25 agents for two days, costing thousands of dollars in token credits and taking multiple days to complete. To enhance real-time interactivity, future work can explore parallelizing agents or developing language models specifically designed for building generative agents. In general, with advances in underlying models, we believe that agents’ performance will improve.
在这项工作中，我们介绍了生成式代理，并展示了其架构的初始实现和评估。未来的研究可以在所提出的代理架构基础上进行改进和进一步评估其性能。在实施方面，例如，检索模块可以通过微调组成检索功能的相关性、新颖性和重要性函数来提高检索更相关信息的能力。此外，可以努力改进架构的性能，使其更具成本效益。本研究需要大量时间和资源来模拟 25 个代理两天，耗费了成千上万美元的代币信用，并需要多天才能完成。为了增强实时互动性，未来的工作可以探索将代理并行化或开发专门设计用于构建生成式代理的语言模型。总的来说，随着基础模型的进步，我们相信代理的性能将会提高。

In terms of evaluation, the assessment of generative agents’ behavior in this study was limited to a relatively short timescale and a baseline human crowdworker condition. While the crowdworker condition provided a helpful comparison point, it did not represent the maximal human performance that could serve as the gold standard in terms of believability. Future research should aim to observe the behavior of generative agents over an extended period to gain a more comprehensive understanding of their capabilities and establish rigorous benchmarks for more effective performance testing. Additionally, varying and contrasting the underlying models, as well as the hyperparameters used for the agents during future simulations, could provide valuable insights into the impact of these factors on the agents’ behavior. Lastly, the robustness of generative agents is still largely unknown. They may be vulnerable to prompt hacking, memory hacking—where a carefully crafted conversation could convince an agent of the existence of a past event that never occurred—and hallucination, among other issues. Future research can comprehensively test these robustness concerns, and as large language models become more resilient to such attacks, generative agents can adopt similar mitigations.
在评估方面，本研究对生成代理的行为评估仅限于相对较短的时间尺度和基准人类众包工作者条件。尽管众包工作者条件提供了一个有用的比较点，但它并不能代表可能作为可信度黄金标准的最大人类表现。未来的研究应该旨在观察生成代理的行为，以获得对其能力的更全面理解，并为更有效的性能测试建立严格的基准。此外，在未来的模拟中变化和对比底层模型以及代理使用的超参数，可能会为这些因素对代理行为的影响提供有价值的见解。最后，生成代理的鲁棒性仍然大部分是未知的。它们可能容易受到提示性黑客攻击、记忆黑客攻击——通过精心设计的对话可能会说服代理相信过去从未发生过的事件的存在——以及幻觉等其他问题的影响。 未来的研究可以全面测试这些鲁棒性问题，随着大型语言模型对此类攻击变得更加弹性，生成式代理可以采用类似的缓解措施。

In general, any imperfections in the underlying large language models will be inherited by generative agents. Given the known biases of language models, generative agents may potentially exhibit biased behavior or stereotypes. Moreover, like many large language models, generative agents may struggle to generate believable behavior for certain subpopulations, particularly marginalized populations, due to limited data availability. While improvements to the agents’ modules may mitigate some of these issues, we believe that addressing them fundamentally requires improving the underlying large language models by aligning their values with the desired outcomes of the agents.
总的来说，基础大型语言模型中的任何不完美都将被生成代理继承。鉴于语言模型的已知偏见，生成代理可能表现出有偏见的行为或刻板印象。此外，像许多大型语言模型一样，生成代理可能会因为数据有限而难以为某些次群体，特别是边缘化人群生成可信的行为。虽然改进代理的模块可能会缓解其中一些问题，但我们认为，从根本上解决这些问题需要通过将基础大型语言模型的价值观与代理的期望结果保持一致。

Generative agents, while offering new possibilities for human-computer interaction, also raise important ethical concerns that must be addressed. One risk is people forming parasocial relationships with generative agents, even when such relationships may not be appropriate. Despite being aware that generative agents are computational entities, users may anthropomorphize them or attach human emotions to them (Hofstadter, 1995; Reeves and Nass, 1996). While this tendency may increase user engagement, it also poses risks, such as users becoming overly reliant on or emotionally attached to the agents (Abercrombie et al., 2023). To mitigate this risk, we propose two principles. First, generative agents should explicitly disclose their nature as computational entities. Second, developers of generative agents must ensure that the agents, or the underlying language models, are value-aligned so that they do not engage in behaviors that would be inappropriate given the context, for example, reciprocating confessions of love.
生成式代理虽然为人机交互提供了新的可能性，但也引发了必须解决的重要伦理问题。一个风险是人们与生成式代理形成偏社交关系，即使这种关系可能不合适。尽管用户意识到生成式代理是计算实体，但他们可能会赋予它们拟人化的特征或将人类情感附加到它们身上。虽然这种倾向可能增加用户参与度，但也存在风险，比如用户过度依赖或情感依恋代理（Abercrombie 等，2023 年）。为了减轻这种风险，我们提出两个原则。首先，生成式代理应明确披露其作为计算实体的本质。其次，生成式代理的开发者必须确保代理或底层语言模型是价值对齐的，以便它们不会参与在特定情境下不合适的行为，例如回应爱的表白。

A second risk is the impact of errors. For example, if a ubiquitous computing application makes the wrong inference about a user’s goals based on generative agent predictions, it could lead to annoyance at best and outright harm at worst. In our instantiation of generative agents, we mitigate these risks by focusing on an interactive video game environment, where such harms are unlikely. However, in other application domains, it will be important to follow best practices in human-AI design (Amershi et al., 2019; Yang et al., 2020) to understand errors and how they might percolate into the user experience.
第二个风险是错误的影响。例如，如果一个无处不在的计算应用程序根据生成代理预测对用户目标做出错误推断，最好会导致恼怒，最坏会导致直接伤害。在我们的生成代理实例中，我们通过专注于交互式视频游戏环境来减轻这些风险，在这种环境中，这样的伤害是不太可能发生的。然而，在其他应用领域中，遵循人工智能设计的最佳实践（Amershi 等，2019 年；Yang 等，2020 年）以了解错误及其可能渗入用户体验将是重要的。

Third, generative agents may exacerbate existing risks associated with generative AI, such as deepfakes, misinformation generation, and tailored persuasion. To mitigate this risk, we suggest that platforms hosting generative agents maintain an audit log of the inputs and generated outputs. This would enable the detection, verification, and intervention against malicious use. While logging alone cannot directly prevent such misuse, it can reduce the likelihood of motivated actors engaging in this behavior, as the risk of disclosure would be higher. Additionally, building this architecture oneself can be time-consuming (in our case, roughly a year), which may deter some actors from pursuing such behavior by using their own generative agent infrastructures.
第三，生成式代理可能会加剧与生成式人工智能相关的现有风险，如深度伪造、信息生成和定制说服。为了减轻这种风险，我们建议托管生成式代理的平台保留输入和生成输出的审计日志。这将使检测、验证和干预恶意使用成为可能。虽然仅仅记录日志不能直接防止这种滥用，但它可以降低有动机的行为者参与此行为的可能性，因为披露的风险会更高。此外，自行构建这种架构可能会耗费大量时间（在我们的情况下大约一年），这可能会阻止一些行为者通过使用自己的生成式代理基础设施来从事这种行为。

A fourth risk is over-reliance: the concern that developers or designers might use generative agents and displace the role of humans and system stakeholders in the design process (Park et al., 2022). We suggest that generative agents should never be a substitute for real human input in studies and design processes. Instead, they should be used to prototype ideas in the early stages of design when gathering participants may be challenging or when testing theories that are difficult or risky to test with real human participants. By adhering to these principles, we can ensure that the deployment of generative agents in the wild is ethical and socially responsible.
第四个风险是过度依赖：担心开发人员或设计师可能使用生成代理并取代人类和系统利益相关者在设计过程中的角色（Park 等，2022 年）。我们建议生成代理永远不应该成为研究和设计过程中真实人类输入的替代品。相反，它们应该在设计的早期阶段用于原型设计想法，当收集参与者可能具有挑战性或测试真实人类参与者难以测试的理论时。遵循这些原则，我们可以确保生成代理在野外部署是符合道德和社会责任的。