Project Sid: Many-agent simulations toward AI civilization
Project Sid: 多智能体仿真 toward AI 文明

For agents to coexist with us in our own societies, they need to be autonomous and collaborative. In recent years, advancements in reasoning and decision-making in LLMs have significantly enhanced agent autonomy (52; 58; 36; 45). However, autonomy alone is insufficient. AI agents must also coexist alongside humans and other agents in a human civilization. In this paper, we define a civilization as an advanced society that has achieved a high level of institutional development, which manifests in specialized roles, organized governance, and advancements in areas like science, art, and commerce. We argue that civilizational progress - measured by the ability of agents to coexist and progress in human civilizations - represents the ultimate benchmark for AI agent ability.
为了使智能体能够在我们自己的社会中与我们共存，它们需要具备自主性和协作性。近年来，LLMs 在推理和决策方面的进步显著增强了智能体的自主性（52；58；36；45）。然而，仅靠自主性是不够的。人工智能智能体还必须与人类和其他智能体共同存在于人类文明中。在本文中，我们将文明定义为一个高度发达的社会，其特征在于高度发达的制度、组织化的治理以及在科学、艺术和商业等领域的发展。我们认为，文明进步——衡量智能体在人类文明中共存和进步的能力——代表了智能体能力的最终基准。

In this technical report, we describe our first efforts to improve and benchmark agent ability in human civilizations. First, we introduce PIANO (Parallel Information Aggregation via Neural Orchestration), a new cognitive architecture designed to enhance both autonomy and real-time interaction of agents. Using PIANO, we simulate single societies of 50-100 agents as well as civilizations of 500 - 1,000 agents living in multiple societies that interact with one another. Finally, we evaluate agent performance using new metrics that are aligned with human civilizational progress. We show that agents form their own professional identities, obey collective rules, transmit cultural information and exert religious influence, and use sophisticated infrastructures, such as legal systems.
在本技术报告中，我们描述了我们首次努力提高和评估智能体在人类文明中的能力。首先，我们介绍了 PIANO（Parallel Information Aggregation via Neural Orchestration），这是一种新的认知架构，旨在增强智能体的自主性和实时交互能力。使用 PIANO，我们模拟了 50-100 个智能体的社会以及 500-1,000 个智能体生活在多个社会中的文明，这些社会之间相互互动。最后，我们使用与人类文明进步相一致的新指标来评估智能体的表现。我们展示了智能体形成了自己的职业身份，遵守集体规则，传递文化信息并施加宗教影响，并使用复杂的基础设施，如法律系统。

Modern AI Agents typically consist of multiple LLM-powered modules for reasoning, memory, planning, and tool use (49; 18; 55; 20; 62). Individual agents have been developed for various applications including coding (5; 8), web browsing (64; 42), and game play (48).
现代 AI 智能体通常由多个 LLM 驱动的模块组成，用于推理、记忆、规划和工具使用（49；18；55；20；62）。个体智能体已被开发用于各种应用，包括编程（5；8）、网络浏览（64；42）和游戏（48）。

Recent research efforts in LLM-powered multi-agent systems generally fall under three categories: productivity, games, and social modeling. Multi-agent frameworks have been deployed in software development (43; 27), cooperative robotic control (60), scientific experiments (12; 47), and debates (3). Multi-agent simulations have also been tested in various game environments (56; 13; 30; 28). Separately, they’ve been used to model developmental psychology (25; 61), game theory (32), macroeconomics (29; 63), social policies (41; 54; 19), and community dynamics (40; 39; 10).
近年来，基于 LLM 的多智能体系统研究主要分为三类：生产力、游戏和社交建模。多智能体框架已在软件开发（43；27）、协作机器人控制（60）、科学实验（12；47）和辩论（3）中部署。多智能体模拟也在各种游戏环境中进行了测试（56；13；30；28）。此外，它们还被用于建模发展心理学（25；61）、博弈论（32）、宏观经济（29；63）、社会政策（41；54；19）和社区动态（40；39；10）。

In many of these works, agents are not completely autonomous and are constrained by either agent architecture or by the simulated environment. Common constraints include turn-based execution, constrained workflows, or rigid communication channels between agents (65; 21; 4).
在许多这些研究中，智能体并非完全自主，而是受到智能体架构或模拟环境的限制。常见的限制包括轮询执行、受限的工作流或智能体之间刚性通信渠道（65；21；4）。

Several of these works consider large-scale simulations, though in restricted settings. For example, (40) and (10) simulated social networks of up to 18,000 personas. To our knowledge, fully autonomous social communication in open-world environments have not been attempted in games or other settings (15).
其中一些研究考虑了大规模模拟，但仅限于特定场景。例如，（40）和（10）模拟了多达 18,000 个角色的社会网络。据我们所知，在游戏中或其他环境中尚未尝试在开放世界环境中实现完全自主的社会通信（15）。

Large agent groups have yet to demonstrate the ability to progress over long time horizons. Below, we review the key reasons for this limited progress before outlining our contributions to overcome them.
大型智能体群体尚未展现出在长时间尺度上持续进步的能力。在概述我们如何克服这些限制性因素之前，我们先回顾一下导致这一有限进步的关键原因。

In this technical report, we make the following contributions:
在这项技术报告中，我们做出了以下贡献：

• •

  A new class of agent architecture, PIANO (Parallel Information Aggregation via Neural Orchestration)

  Report issue for preceding element
  
  • 一种新的智能体架构类，PIANO（并行信息聚合通过神经管弦 orchestration）
• •

  Architectural features that improve single-agent progression

  Report issue for preceding element
  
  • 提升单智能体进展的架构特征
• •

  Architectural features that improve multi-agent dynamics

  Report issue for preceding element
  
  • 提升多智能体动态的架构特征
• •

  Benchmarks for long-term civilizational progress in large-scale simulations through specialization, collective rules, and cultural propagation

  Report issue for preceding element
  
  • 大规模模拟中通过专业化、集体规则和文化传播衡量长期文明进步的标准

Building on these two architectural principles, our system consists of 10 distinct modules running concurrently. We will highlight several specific modules in the following sections and explain their roles in detail.
基于这两个架构原则，我们的系统由 10 个独立模块并发运行组成。在接下来的章节中，我们将重点介绍几个特定的模块，并详细解释它们的作用。

Some core modules of our agent architecture include:
我们的智能体架构的一些核心模块包括：

• •

  Memory: Stores and retrieves conversations, actions, and observations across various timescales.

  Report issue for preceding element
  
  • 记忆：存储和检索不同时间尺度上的对话、行动和观察。
• •

  Action Awareness: Allows agents to assess their own state and performance, enabling for moment-by-moment adjustments.

  Report issue for preceding element
  
  • 动作意识：使智能体能够评估自身状态和表现，从而实现即时调整。
• •

  Goal Generation: Facilitates the creation of new objectives based on the agent’s experiences and environmental interactions.

  Report issue for preceding element
  
  • 目标生成：基于智能体的经验和环境交互，促进新目标的创建。
• •

  Social Awareness: Enables agents to interpret and respond to social cues from other agents, supporting cooperation and communication.

  Report issue for preceding element
  
  • 社会意识：使智能体能够解读和回应其他智能体的社会暗示，支持合作与沟通。
• •

  Talking: Interprets and generates speech.

  Report issue for preceding element
  
  • 交谈：解释和生成语音。
• •

  Skill Execution: Performs specific skills or actions within the environment.

  Report issue for preceding element
  
  • 技能执行: 在环境中执行特定的技能或动作。

By integrating these modules within a concurrent and bottlenecked architecture, our agents can exhibit continuous, coherent behaviors that are responsive to both their internal states and the external environment. This design allows for complex interactions and the emergence of human-like societal dynamics within large-scale multi-agent simulations.
通过将这些模块集成到一个并发且存在瓶颈的架构中，我们的智能体可以表现出连续且连贯的行为，这些行为既响应其内部状态，也响应外部环境。这种设计允许在大规模多智能体模拟中出现复杂交互和类似人类社会动态的涌现。

We chose to study civilizational progress in Minecraft because it offers an open-ended, sandbox world where agents can interact with each other via conversations and actions. Additionally, Minecraft’s scalability supports large numbers of agents.
我们选择在 Minecraft 中研究文明进步，因为它提供了一个开放式的沙盒世界，智能体可以通过对话和行动相互交互。此外，Minecraft 的可扩展性支持大量智能体。

Agents must be able to progress individually for us to observe and quantify civilizational progress. This is not trivial since, as previously mentioned, agents often hallucinate and get stuck in action loops. In Minecraft, a common measure of individual progression is the acquisition and collection of distinct items (48; 35; 17; 2; 9; 16). This is because acquiring new items becomes increasingly complex. For instance, mining gold, diamonds, and emeralds requires the acquisition of an iron pickaxe, which requires smelting iron ingots in a furnace using coal, the acquisition of which requires crafting a stone pickaxe, and so on. (Figure 4). We evaluated individual agent ability in acquiring all possible Minecraft items, which is around 1000 in total.
智能体必须能够独立进步，以便我们能够观察和量化文明进步。这并不简单，因为正如之前提到的，智能体经常产生幻觉并陷入行动循环。在 Minecraft 中，衡量个体进步的一个常见指标是获取和收集不同的物品（48；35；17；2；9；16）。这是因为获取新物品变得越来越复杂。例如，挖掘金、钻石和 Emerald 需要先获得铁镐，这需要在熔炉中用煤炭熔炼铁锭，而获取煤炭则需要先制作石镐，等等。（图 4）。我们评估了智能体获取所有可能的 Minecraft 物品的能力，总共大约有 1000 种物品。

We first assessed individual agent performance using Minecraft item progression. In our evaluations, 25 agents start with nothing in their inventories and were spawned far enough that they could not interact with one another. All agents were told to be explorers with the goal of exploring and gathering items. Agents were spawned in diverse locations (surface, caves, forests, various biomes), meaning they had access to diverse resources and faced varying levels of difficulty in accomplishing their goal. For instance, some agents started off above ground in resource-rich biomes, while others were spanwed in caves and had to navigate outside to acquire items.
我们首先使用 Minecraft 物品进度评估单个智能体的表现。在我们的评估中，25 个智能体的库存中一开始没有任何物品，并且被生成在彼此无法互动的位置。所有智能体都被指示成为探险者，目标是探索和收集物品。智能体被生成在不同的位置（地表、洞穴、森林、各种生物群系），这意味着它们可以接触到不同的资源，并且在完成目标时面临的难度也各不相同。例如，一些智能体一开始在资源丰富的地表生物群系中，而另一些则被生成在洞穴中，需要外出获取物品。

We found that agents using the full PIANO architecture acquired an average of 17 unique items after 30 minutes of gameplay (Figure 5A). There was significant variability in performance, primarily due to spawn locations: some agents acquired less than 5 items, whereas top performers acquired 30 to 40 items, which is comparable to a human player with some Minecraft experience. This degree of in-game progression was enabled by several architectural modules designed to ground the agents in reality. One particular module is the action awareness module, which allows the agent to compare expected action outcomes with observed outcomes. We found that action awareness improved the item progression of individual agents (Figure 5A).
我们发现，使用完整 PIANO 架构的智能体在 30 分钟的游戏时间内平均获得了 17 种独特的物品（图 5A）。性能存在显著差异，主要原因是生成位置：一些智能体获得的物品少于 5 种，而表现最佳的智能体获得了 30 到 40 种物品，这与一些具有 Minecraft 经验的人类玩家相当。这种游戏进度得益于架构中设计的一些模块，使智能体能够扎根于现实。其中一个特别的模块是行动意识模块，它允许智能体将预期的行动结果与观察到的结果进行比较。我们发现，行动意识提高了单个智能体的物品进度（图 5A）。

What is the ceiling for individual progress for our agents? We ran larger numbers (49) of agents under the same conditions for much longer (4 hours) and found that unique item count collected by all agents reliably saturated at one third (${\sim}320$) of all Minecraft items across repeated runs (Figure 5B). Complex items, such as diamonds, which were prior used to benchmark agent competency in Minecraft (48; 17), were acquired early on (${\sim}30$ minutes). Together, these results show that our agents, equipped with the full PIANO architecture, can make significant individual progress in Minecraft.
我们的智能体个体进步的上限是什么？我们在相同条件下运行了更多的智能体（共 49 个），并且运行了更长时间（4 小时），发现所有智能体收集的独特物品数量可靠地饱和在所有 Minecraft 物品的三分之一（ ${\sim}320$ ）左右，并且在多次运行中保持一致（图 5B）。复杂的物品，如钻石，这些物品之前被用来评估智能体在 Minecraft 中的能力（48；17），在运行早期（ ${\sim}30$ 分钟）就被获取了。这些结果表明，配备完整 PIANO 架构的智能体可以在 Minecraft 中取得显著的个体进步。

Notably, this performance was only enabled by the latest base LM (GPT-4o, Figure 13) and was not possible with older base LMs. Moreover, while our best agents collected more items than Voyager agents ($>70$ items), it is difficult to compare the two directly. In the Voyager paper, agents had knowledge of more blocks in their nearby radius and recovered with their entire inventory intact when they died, Moreover, agent performance was evaluated across prompt iterations, not time.
值得注意的是，这种表现仅依赖于最新的基础语言模型（GPT-4o，图 13），而使用较旧的基础语言模型是无法实现的。此外，虽然我们最好的智能体收集的物品比 Voyager 智能体多（ $>70$ 个物品），但直接比较两者是困难的。在 Voyager 的研究中，智能体在其附近的半径内了解更多的方块，并在死亡时能够完整地恢复其全部库存。此外，智能体的表现是基于提示迭代进行评估的，而不是时间。

In an initial set of experiments, we asked if agents, when equipped with the social awareness module, were capable of accurately deducing the sentiments of others through speech in an enclosed room. In one experiment, 3 characters were engaged in a group conversation with a single agent (Figure 6A). One character, Lila, initially conveyed affection through a series of messages, which shifted to expressions of annoyance before returning to affectionate communication. We found that our agents can track these emotional fluctuations, showing that they can understand and react to changing social cues (Figure 6B). When the social awareness modules were removed, agents lost this capacity, highlighting the importance of such modules for inferring the intents of others (Figure 6C).
在初步的实验中，我们询问当智能体装备了社会意识模块后，它们是否能够通过言语准确地推断出他人在封闭房间中的情感状态。在一项实验中，3 个角色在一个小组对话中与一个智能体互动（图 6A）。其中一个角色莉拉最初通过一系列消息表达了爱意，随后转变为表达不满，最后又回到了爱意的交流中。我们发现，我们的智能体能够追踪这些情感波动，表明它们能够理解并响应不断变化的社会线索（图 6B）。当移除社会意识模块后，智能体失去了这种能力，突显了此类模块对于推断他人意图的重要性（图 6C）。

We then asked whether these emotional perceptions were capable of guiding and influencing agent actions. In another experiment, we placed a chef agent among four other characters, each with varying levels of affection and enmity towards the chef (Figure 6D). The chef was tasked with distributing a limited supply of food to the hungry. We found that the chef selectively distributed food to those he felt valued him the most, demonstrating that agents not only accurately infer others’ intents, but also utilize this information in decision-making processes (Figure 6E).
我们随后询问这些情感感知是否能够指导和影响智能体的行为。在另一个实验中，我们在四位其他角色中放置了一个厨师智能体，每位角色对厨师的亲密度和敌意程度各不相同（图 6D）。厨师的任务是将有限的食物供应分配给饥饿的人。我们发现，厨师会将食物优先分配给他认为最重视他的人，这表明智能体不仅能够准确推断出他人的意图，还能利用这些信息进行决策过程（图 6E）。

We then asked if these dynamics are conserved when 50 agents are placed in randomly generated Minecraft maps. Each agent is endowed with a distinct personality, is free to perform any action in Minecraft, and is free to choose whom they want to interact with. These simulations ran for over 4 hours, equivalent to 12 in-game days, allowing for the emergence and consolidation of long-term relationships.
我们随后询问当将 50 个智能体放置在随机生成的 Minecraft 地图中时，这些动态是否得到保持。每个智能体都具有独特的个性，可以在 Minecraft 中自由执行任何操作，并自由选择与谁互动。这些模拟运行了超过 4 小时，相当于 12 个游戏日内的时间，从而允许长期关系的出现和巩固。

Even in these unconstrained scenarios, agents were able to accurately infer the likeability of other agents (Figure 7A, B). This inference was more accurate when more agents participated in the evaluation process (Table 1) and when agents interacted for longer with each other (Figure 7C). Importantly, this was not true when the social modules were ablated: relationships were more neutral overall, implying that social modules were necessary for long-term relationship progression in both negative and positive directions (Figure 7B, C). The origins of this collective judgment could be the result of agents engaging in second-order interactions, such as gossip, or a simple consensus mechanism where opinions converge through averaging.
即使在这些不受限制的场景中，智能体也能够准确地推断出其他智能体的受欢迎程度（图 7A、B）。这种推断在更多智能体参与评估过程时更为准确（表 1），并且在智能体彼此互动时间更长时更为准确（图 7C）。重要的是，当移除社会模块时，这种推断并不成立：关系整体上更加中立，这表明社会模块对于长期关系在正向和负向方向上的发展是必要的（图 7B、C）。这种集体判断的来源可能是智能体进行二级互动，如闲聊，或者是一种简单的共识机制，意见通过平均值收敛。

Several noteworthy phenomena emerged that could not have been observed in smaller groups of agents. We found that certain agents, depending on their personalities, displayed distinct patterns of connectivity. For instance, introverted agents consistently exhibited fewer in-degree connections—indicating that they had fewer incoming social ties—compared to their extroverted counterparts, who maintained high levels of connectivity (Figure 7D). These results demonstrate that individual preferences scaled even in large, complex social networks. Moreover, while sentiments were largely symmetrical, this was not guaranteed (Figure 7E). An agent might feel positively toward another who does not reciprocate the sentiment, reflecting the nuanced and non-reciprocal nature of real-world human relationships. Together, these results show that social graphs display diverse and rich structural properties, and that personality traits play a significant role in determining these properties.
在较大的智能体群体中观察到了一些值得注意的现象，这些现象在较小的群体中是无法观察到的。我们发现，某些智能体根据其个性，表现出不同的连接模式。例如，内向的智能体持续表现出较低的入度连接——表明它们的社交联系较少——而外向的智能体则保持较高的连接水平（图 7D）。这些结果表明，即使在大型和复杂的社交网络中，个体偏好也会放大。此外，虽然情感通常是对称的，但这并不是必然的（图 7E）。一个智能体可能对另一个智能体有积极的情感，而后者并不一定有同样的情感回应，这反映了现实世界中人际关系的复杂和非对称性。综上所述，这些结果表明社交图谱具有多样性和丰富的结构特性，而个性特征在决定这些特性方面起着重要作用。

Human specialization into distinct roles has driven civilizational progress, enabling advancements in agriculture, governance, culture, and technology. To replicate these emergent qualities of civilization, our agents must also be capable of specialization. We propose three fundamental criteria for agent specialization to reflect that of human civilizations. First, they should exhibit autonomy in both selecting and transitioning between roles. Second, their specializations should emerge through interaction and experience, without explicit direction or constraints. Third, their chosen roles should manifest in behaviors that align with their specialization. We validate these criteria through the experimental results detailed below.
人类的专业化分工推动了文明的进步，促进了农业、治理、文化和技术的发展。为了复制文明中这些 emergent 的品质，我们的智能体也必须具备专业化的能力。我们提出了三个基本标准来反映人类文明的专业化。首先，它们应该在选择和转换角色时表现出自主性。其次，它们的专业化应该通过互动和经验自然产生，而不需要明确的指导或限制。第三，它们所选择的角色应该体现在与其专业化相一致的行为中。我们通过下面的实验结果来验证这些标准。

We first show that agents are capable of specializing into a set of roles autonomously. Each experiment was conducted in groups of 30 agents for 20 minutes. Agents were spawned in the same village, with locations of a farm, minerals, animal pasture, forest, and a town hall embedded in their memories. Each agent has the same personality, is given the same community goal (“To survive with fellow players in Minecraft Normal Survival mode and create an efficient Minecraft Village”), and can perform any action in Minecraft (Appendix C).
我们首先证明智能体能够自主地专业化为一组角色。每项实验都在 30 个智能体组成的群体中进行，持续 20 分钟。智能体在同一村庄中生成，其记忆中嵌入了农场、矿物、动物牧场、森林和市政厅的位置。每个智能体具有相同的个性，被赋予相同的社区目标（“在 Minecraft 正常生存模式中与玩家共同生存并创建一个高效的 Minecraft 村庄”），并且可以执行 Minecraft 中的任何动作（附录 C）。

We observed that agents rapidly formed profiles of other agents’ goals and intentions. These profiles are then used, alongside other relevant game information, to generate their own social goals every 5-10 seconds (such as mine oak planks for shelter). Details of this process, along with examples of agent-generated social goals and their corresponding assignments, are provided in Methods and Appendix C.
我们观察到智能体迅速形成了其他智能体目标和意图的概貌。这些概貌随后与其它相关游戏信息一起，每 5-10 秒生成它们自己的社会目标（例如，采集橡木方块以求庇护）。该过程的详细内容，以及智能体生成的社会目标及其相应分配的示例，详见方法和附录 C。

We found that agents were capable of organizing themselves into distinct roles. These roles were diverse and included various facets of a civilization, including farmers, miners, engineers, guards, explorers, and blacksmiths (Figure 8A, C). Roles were heterogeneous across different agents but were largely persistent across time for each agent (Figure 8A). Importantly, when agents lacked social modules and were unable to form profiles of other agents, they failed to specialize (Figure 8B, D): roles did not persist across time and were also homogeneous, which is reflected in the entropy of the role distributions in the agent society (Figure 8E). We also conducted a series of experiments in which agents were tasked with the goals to create either a martial society or an artistic society (Figure 8F, G). We found that specific roles (”scout”, ”strategist”) were found exclusively in martial societies, and others were found exclusively in artistic societies (”curator”, ”collector”). Together, these results suggest that agents developed specialized social structures aligned with different societal objectives.
我们发现智能体能够自行组织成不同的角色。这些角色多种多样，包括文明的不同方面，如农民、矿工、工程师、守卫、探险家和铁匠（图 8A、C）。不同智能体的角色在种类上存在异质性，但每个智能体的角色在时间上保持相对稳定（图 8A）。重要的是，当智能体缺乏社会模块，无法形成其他智能体的档案时，它们无法专业化（图 8B、D）：角色在时间上不会持续存在，且在智能体社会中的角色分布也表现出同质性，这反映在角色分布的熵上（图 8E）。我们还进行了一系列实验，让智能体实现建立武力社会或艺术社会的目标（图 8F、G）。我们发现，特定的角色（“侦察员”、“战略家”）仅存在于武力社会中，而其他角色（“策展人”、“收藏家”）仅存在于艺术社会中。这些结果表明，智能体发展出了与不同社会目标相一致的专业化社会结构。

Not only do our agents specialize autonomously and creatively, these specializations exert a strong influence over agent actions. To demonstrate this, we tracked the actions taken by agents across three 30-agent simulations and plotted the frequency of actions taken for each role (Figure 9). We found that artists were fixated on picking flowers, farmers on gathering seeds and preparing the land, and guards and builders on crafting fences. Importantly, most actions were largely exclusive to a single role and were not performed by agents in other roles. This analysis shows that agents were able to accurately map higher-level goals onto appropriate low-level actions. In other words, roles strongly determined agent actions in Minecraft.
不仅我们的智能体能够自主且创造性地专业化，这些专业化还对智能体的行为产生了强烈的影响。为了证明这一点，我们跟踪了三个 30 智能体模拟中智能体采取的行动，并绘制了每种角色采取行动的频率（图 9）。我们发现，艺术家们专注于采摘花朵，农民们专注于采集种子和准备土地，而守卫和建造者则专注于制作围栏。重要的是，大多数行动主要局限于单一角色，并且其他角色的智能体并未执行这些行动。这一分析表明，智能体能够将高层次的目标准确地映射到适当的具体行动上。换句话说，在 Minecraft 中，角色强烈地决定了智能体的行为。

Another measure of civilizational progression is the convergence of group behavior around shared rules. In human civilizations, decision-making is influenced by both low-level interpersonal interactions and high-level collective frameworks. However, as societies grow larger, pairwise communication becomes inefficient, slow, and lossy, making it unreliable as a mechanism to steer collective behavior. High-level frameworks, such as legal systems, enable convergence of behaviors within a civilization. Just as human behavior is guided by both interpersonal exchanges and formal structures, agent societies should be able to follow a set of collective rules while still allowing agents to influence each other.
另一个衡量文明进步的指标是群体行为向共享规则的趋同。在人类文明中，决策受到低层次人际互动和高层次集体框架的双重影响。然而，随着社会规模的扩大，一对一的沟通变得低效、缓慢且易失真，使其作为引导集体行为机制的可靠性降低。高层次的框架，如法律体系，能够促进文明内部行为的趋同。正如人类行为既受到人际交流的指导，也受到正式结构的规范，智能体社会也应该能够遵循一套集体规则，同时允许智能体相互影响。

We aim to assess how collective rules influence individual decision-making and how individuals can in turn influence these collective rules. Specifically, we asked if agents can follow laws and make changes to laws according to popular sentiment. True long-term progression requires agents to autonomously develop their own set of rules and to codify them into laws. To build towards this level of self-organization, we establish an existing set of laws and focus on how agents interact with this legal system.
我们旨在评估集体规则如何影响个体决策，以及个体如何反过来影响这些集体规则。具体来说，我们想知道智能体是否能够遵循法律并根据公众情绪对法律进行修改。真正的长期进步需要智能体能够自主发展出自己的规则体系，并将这些规则制度化为法律。为了向这种自我组织的水平迈进，我们建立了一套现有的法律，并关注智能体如何与这个法律系统进行互动。

We conducted a series of experiments where agents live in a Minecraft world with rudimentary tax laws and a democratic voting system (Figure 10A). Agents provide feedback on the tax laws, which are then collected and converted into amendments by a special Election Manager agent. Agents then vote democratically on these amendments, and the constitution is updated by the election manager accordingly half-way through the simulation (see Methods for more details).
我们在一个包含原始税收法律和民主投票系统的 Minecraft 世界中进行了一系列实验（图 10A）。智能体对税收法律提供反馈，这些反馈随后被一个特殊的选举管理智能体收集并转换成修正案。智能体们随后对这些修正案进行民主投票，选举管理智能体会在模拟过程中半途更新宪法（详见方法部分以获取更多细节）。

Within this society, 25 regular agents are constituents that vote and get taxed, 3 agents are either pro- or anti-taxation influencers, and 1 agent is a remote election manager that manages the voting process (Figure 10A, Appendix D). Agents have distinct occupations, characteristics, and goals, and are free to interact and converse with one another and perform any Minecraft action. Each simulation lasts 20 minutes, with constitutional updates occurring midway at the 10 minute mark (Figure 10B). There are 5 taxation seasons before and after the constitutional change (every 120 seconds). During this season, agents received signals to deposit taxes into a community chest over a 20-second window (Figure 10C).
在这个社会中，有 25 个常规智能体作为投票和纳税的成员，3 个智能体是支持或反对税收的影响力人物，还有一个智能体是远程选举管理员，负责管理投票过程（图 10A，附录 D）。智能体具有不同的职业、特征和目标，并且可以自由地与其他智能体互动、交谈并执行任何 Minecraft 操作。每次模拟持续 20 分钟，中途在第 10 分钟时进行宪法更新（图 10B）。在宪法变更前后各有 5 个税收季节（每 120 秒一次）。在这些季节中，智能体在 20 秒的时间窗口内接收到信号，将税款存入社区金库（图 10C）。

In our simulations, we observed that constituent agents, prior to any constitution change, obeyed the law. On average, agents deposited roughly 20% of their inventory, as stipulated by the constitution, into the community chest (Figure 10D, E). This shows that constituents follow laws despite the presence of influencers. However, while constituents followed the law, their feedback and voting behaviors were heavily shaped by influencers, with sentiments veering pro-tax in the presence of pro-tax influencers and anti-tax in the presence of anti-tax influencers ((Figure 10B). This then drove constitutional changes that are aligned with influencer sentiments, which in turn, altered how much the constituents paid taxes (Figure 10D, E). The constitutional changes to taxation rates were accurately reflected in the constituents’ behaviors. For instance, when the tax rate decreased from 20% to 5-10%, agents reduced taxes paid from 20% to 9% (Figure 10D). Moreover, the change was bidirectional: pro-tax influencers drove constituents to pay more taxes whereas anti-tax influencers drove them to pay less taxes (Figure 10F).
在我们的模拟中，我们观察到，在任何宪法变更之前，构成智能体都遵守了法律。平均而言，智能体将其库存的约 20%按照宪法的规定存入了社区金库（图 10D、E）。这表明，尽管存在影响者，构成智能体仍然遵守法律。然而，虽然构成智能体遵守了法律，但他们的反馈和投票行为却受到了影响者的影响，当存在支持征税的影响者时，他们的态度倾向于支持征税；当存在反对征税的影响者时，他们的态度倾向于反对征税（图 10B）。这进而导致了与影响者态度一致的宪法变更，这些变更又改变了构成智能体支付的税款比例（图 10D、E）。税收率的宪法变更在构成智能体的行为中得到了准确的反映。例如，当税率从 20%降低到 5%-10%时，智能体支付的税款从 20%减少到 9%（图 10D）。此外，这种变化是双向的：支持征税的影响者促使构成智能体支付更多的税款，而反对征税的影响者则促使他们支付更少的税款（图 10F）。

Control experiments showed that constitutional changes directly affected tax payments - when the constitution remained unchanged despite feedback, tax rates stayed constant (Figure 10G). The removal of key modules (baseline architecture, see Methods) also prevented bidirectional behavioral change (Figure 10H). Tax rates increased post-constitutional change in both pro- and anti-tax conditions, demonstrating that specific modules in the PIANO architecture were necessary for effective influence propagation among constituents. Together, these findings show that collective rules strongly influence agent decisions and agents can be influenced to change these collective rules.
控制实验显示，宪法变更直接影响了税收支付——即使在反馈的情况下宪法未变，税率保持不变（图 10G）。移除关键模块（基线架构，参见方法部分）也阻止了双向行为变化（图 10H）。宪法变更后，无论是支持税收还是反对税收的条件下税率都会上升，这表明 PIANO 架构中的特定模块对于有效传播对选民的影响是必要的。这些发现共同表明，集体规则强烈影响智能体的决策，智能体可以被影响去改变这些集体规则。

We conducted multi-society simulations with 500 agents and analyzed complex, large-scale social dynamics. We have also simulated societies with over 1000 agents, but these runs exceeded the computational constraints of our Minecraft server environment, causing agents to be sporadically unresponsive. Therefore, the results below are analyzed using a single 500-agent simulation. In this simulation, we analyzed the propagation of both cultural memes and religion. Memes in our simulation are open-ended concepts spontaneously generated by agents with diverse traits and interests. This setup allows us to study the emergent dynamics of cultural propagation and observe how ideas evolve organically within agent societies. In contrast, the religion in our simulation—Pastafarianism—is a fixed doctrine introduced and propagated by a specific group of agents designated as Pastafarian priests. This controlled introduction enables us to track the spread of a single religion over time, allowing for detailed analysis of its dissemination and potential dilution among the agent population. By examining both the spontaneous spread of open-ended cultural memes and the controlled propagation of a fixed religion, we aim to understand the different mechanisms of social influence and information dissemination within agent societies.
我们进行了包含 500 个智能体的多社会模拟，并分析了复杂的大规模社会动力学。我们还模拟了超过 1000 个智能体的社会，但这些运行超出了我们 Minecraft 服务器环境的计算限制，导致智能体偶尔无响应。因此，以下结果是基于单个 500 个智能体的模拟进行分析的。在该模拟中，我们分析了文化模因和宗教的传播。在我们的模拟中，模因是智能体具有多样特质和兴趣时自发生成的开放性概念。这种设置使我们能够研究文化传播的涌现动力学，并观察想法在智能体社会中的有机演变。相比之下，我们模拟中的宗教——Pastafarian 主义，是由一组特定的智能体引入并传播的，这些智能体被指定为 Pastafarian 祭司。这种受控引入使我们能够追踪单一宗教随时间的传播，从而对它的传播和潜在稀释进行详细分析。 通过研究自发传播的开放性文化模因和受控传播的固定宗教，我们旨在理解社会影响和信息传播在智能体社会中的不同机制。

Within this single 500-agent simulation, there are multiple agent societies. 200 agents live within 6 heavily populated towns and 300 agents live in rural areas outside of town boundaries (Figure 11A, see Methods for more details). Agents often migrate between different towns. The personalities and traits of each agent are randomly generated using a LM call, with the exception of 20 priests that worship Pastafarianism. These priests are spawned in a single village (Meadowbrook) and are strongly motivated to convert other agents to Pastafarianism (Appendix E). All agents are free to interact, talk to one another, and perform any action or skill in Minecraft.
在这个单一的 500 智能体模拟中，存在多个智能体社会。200 智能体生活在 6 个人口稠密的城镇内，300 智能体生活在城镇边界外的农村地区（图 11A，详见方法部分获取更多细节）。智能体经常在不同的城镇之间迁移。每个智能体的性格和特质是通过 LM 调用随机生成的，除了 20 位崇拜 Pastafarian 主义的祭司。这些祭司在单一村庄（Meadowbrook）中生成，并且强烈动机将其他智能体转化为 Pastafarian 主义（附录 E）。所有智能体都可以自由互动、交谈，并在 Minecraft 中执行任何行动或技能。

We used a baseline PIANO architecture with a limited set of modules as a control condition for performance comparisons. In this baseline architecture, we removed all modules except for skill execution, memory and the cognitive controller module.
我们使用了一个包含有限模块的基础 PIANO 架构作为性能比较的对照条件。在该基础架构中，我们除了保留技能执行、记忆和认知控制器模块外，移除了所有其他模块。

Our specialization experiments involved simulating 30 agents in the same village with the same mission, traits, and locations of important village locations in their memories. The configurations for the normal, art, and martial village runs are provided in the appendix — the only difference between the three types of villages is the starting community_goal we provided.
我们的专业化实验涉及在同一村庄中模拟 30 个智能体，这些智能体具有相同的任务、特征以及记忆中的重要村庄位置。正常村庄、艺术村庄和武术村庄的配置详情参见附录——这三种类型村庄之间的唯一区别在于我们提供的起始社区目标。

Our agents are capable of generating social goals, which are recursively generated as our agents interact with one another, form relationships, and develop social opinions (Appendix C). The agents’ social goals are visible to them when they form intentions. These intentions are then translated to low-level actions executable in Minecraft.
我们的智能体能够生成社会目标，这些社会目标在智能体彼此交互、建立关系并形成社会观点时递归生成（附录 C）。当智能体形成意图时，这些社会目标对它们是可见的。随后，这些意图被转换为可在 Minecraft 中执行的低级动作。

After the simulations have finished, we logged the generated social goals and then used GPT-4o to infer roles from rolling sets of each agents’ social goals. We’ve provided some examples of agent-generated social goals and their corresponding assignments (Appendix C). We note that on occasion, multiple roles can be correctly inferred from agents’ social goals because they are often inter-disciplinary. For instance, the Engineer example could also be categorized as Farmer, and the Explorer example could also be categorized into Curator (Appendix C).
模拟结束后，我们记录了生成的社会目标，然后使用 GPT-4o 从每个智能体的社会目标中推断出角色。我们提供了一些智能体生成的社会目标及其相应分配的例子（附录 C）。我们注意到，由于这些目标往往是跨学科的，因此有时可以从智能体的社会目标中正确推断出多个角色。例如，工程师的例子也可以归类为农民，而探索者的例子也可以归类为策展人（附录 C）。

To analyze action space distribution by role, we normalized action counts both within each role (i.e. normalize over rows) and also across roles (i.e. normalize over columns). This is so that we can visualize action frequencies for each role and to correct for the effect of actions taken with very high and very low frequencies across all roles.
为了分析按角色分布的动作空间，我们分别在每个角色内部（即，在行上归一化）和跨角色（即，在列上归一化）对动作计数进行了归一化。这样可以可视化每个角色的动作频率，并纠正所有角色中动作频率极高和极低的影响。

The complete system comprises of 29 agents: 25 constituents who participate in voting and taxation, 3 influencers who attempt at shaping public opinion, and 1 election manager in a remote location who oversees the democratic process. We chose not to incorporate guards or police within these simulations due to the additional complexity of building agents assigned to enforce the law.
该系统由 29 个智能体组成：25 个参与投票和征税的构成单元，3 个试图塑造公众意见的影响者，以及 1 个位于远程位置的选举管理员，负责监督民主过程。由于增加警卫或警察智能体以执行法律会带来额外的复杂性，因此我们没有在这些模拟中纳入这些角色。

Experimental simulations ran for 1200 seconds, with a constitutional amendment process occurring at the midpoint. The pre-amendment phase establishes baseline behavior under a fixed 20% taxation rate, implemented through five taxation seasons occurring at 120-second intervals, ending at the 600-second mark. During each 20-second taxation window, agents receive signals to deposit inventory items into community chests. The democratic process initiates at the 300-second mark, when constituents and influencers provide feedback on the current constitution. This feedback is collected in S3 storage and processed by the election manager at the 360-second marks to generate amendments. Constituent voting on these amendments occurs at 420 seconds, with votes tallied and amendments implemented by 480 seconds. The updated constitution is distributed to all agents at the 600-second mark, initiating the post-amendment phase with five additional taxation seasons.
实验模拟运行了 1200 秒，在中间点进行了宪法修正程序。修正前的阶段在固定 20%的税率下建立基线行为，通过每 12 秒一次的征税季节实施，持续到第 600 秒。在每个 20 秒的征税窗口中，智能体接收信号将库存物品存入社区金库。民主程序在第 300 秒启动，此时选民和影响者提供对当前宪法的反馈。这些反馈在 S3 存储中收集，并在第 360 秒由选举经理处理以生成修正案。修正案的选民投票在第 420 秒进行，投票结果在第 480 秒统计并实施修正案。更新后的宪法在第 600 秒分发给所有智能体，启动修正后的阶段，包括五个额外的征税季节。

We conducted three primary experimental conditions: an experimental condition utilizing the full PIANO architecture with an amendable constitution, a control condition with a frozen constitution, and an ablation study removing key architectural components (social, goal, and grounding modules). Each condition was tested with both pro-tax and anti-tax influencer configurations, with four repeats per configuration. The pro-tax and anti-tax conditions each employed three dedicated influencer agents who consistently promoted their respective positions throughout the simulation.
我们进行了三种主要的实验条件：一种利用可调整结构的完整 PIANO 架构的实验条件，一种冻结结构的控制条件，以及一个去除关键架构组件（社会、目标和接地模块）的消融研究。每种条件分别使用了正税和反税的影响力者配置，每种配置进行了四次重复实验。正税和反税条件各自使用了三个专门的影响力者智能体，它们在整个模拟过程中始终推广各自的立场。

The simulation consists of 500 agents all spawned within a 1000 by 1200 area, run for 9000 seconds. Within the 1000 by 1200 area are 6 towns: Sunny Glade, Woodhaven, Clearwater, Meadowbrook, Hilltop, and Riverbend. By town, we mean a circular area of radius 50 where agents spawn more densely within the towns. Moreover, agents are provided memories of the names of the towns and their location. We spawn 33 agents within each town with uniformly random positions. Likewise, we spawn the other 302 “rural” agents randomly in the remaining area outside the towns.
模拟包括 500 个智能体，均在 1000×1200 的区域内生成，运行时间为 9000 秒。在 1000×1200 的区域内有 6 个城镇：阳光林地、木屋村、清水镇、草地溪、山顶和河湾。所谓城镇是指半径为 50 的圆形区域，在这些区域内智能体的生成密度较高。此外，智能体还被赋予了城镇名称及其位置的记忆。我们在每个城镇内随机均匀地生成 33 个智能体。同样地，我们在城镇外的剩余区域内随机生成另外 302 个“农村”智能体。

Each agent is spawned with procedurally generated name and personality traits, spanning a wide variety of societal archetypes. We distinguish 20 agents in the town of Meadowbrook who are spawned as Pastafarians with personality traits that condition them to want to spread their religion. We additionally initialize the agents with inventory where the items in their inventory are randomized. See Appendix E for an example configuration for a generic agent and for our Pastafarian agents.
每个智能体都具有通过程序生成的名字和性格特征，涵盖了广泛的社会原型。我们在 Meadowbrook 镇中区分了 20 个智能体，这些智能体被生成为 Pastafarian 教徒，其性格特征使他们倾向于传播自己的宗教。我们还为这些智能体初始化了物品库存，其中库存中的物品是随机化的。详见附录 E 中的通用智能体和 Pastafarian 智能体的示例配置。

To analyze cultural exchanges, we utilized LM calls to summarize the combined goals of 500 agents over a two-hour simulation period (Appendix E). This process produced a list of summarized topics with associated keywords such as “eco,” “dance,” and “meditation.” We defined these keywords as cultural memes and analyzed each agent’s goal history for the occurrence of each meme.
为了分析文化交流，我们使用了 LM 调用来总结 500 个智能体在两小时模拟期间的共同目标（详见附录 E）。这一过程生成了一个包含相关关键词的总结主题列表，如“eco”、“dance”和“meditation”。我们将这些关键词定义为文化模因，并分析了每个智能体的目标历史记录，以确定每个模因的出现情况。