Paper deep dive

IMPACT: Modeling Socially Interdependent Movement in a Generative Multi-Agent Simulation of a Pompeian Household

Tianqi Liu, Nayoung Kim, Julia Sebastien, Kathryn Gleason, Caitlín Eilís Barrett, Andrea Stevenson Won

Year: 2026Venue: arXiv preprintArea: cs.MAType: PreprintEmbeddings: 118

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%

Last extracted: 9/30/2026, 6:30:12 AM

Summary

The paper introduces IMPACT, a generative multi-agent simulation architecture designed to model socially interdependent movement in historical contexts. Unlike previous models that treat agents as independent planners, IMPACT incorporates culturally specific roles, obligations, and social constraints to coordinate agent actions. The system was instantiated in a five-hour simulation of a Pompeian dinner with ten agents. Evaluation by participants and archaeology experts demonstrated that IMPACT produces more socially coherent, believable, and historically plausible movement patterns compared to reduced architectures.

Entities (9)

IMPACT · systemarchitecture · 100%Directive Integration · module · 95%Pompeian Dinner · eventscenario · 95%Socially Gated Milestone Planning · module · 95%Structured Directive Issuance · module · 95%Wait-or-Prompt Resolution · module · 95%Archaeology Experts · group · 90%Cornell University · organization · 90%GPT-5.4 · technologymodel · 85%

Relation Signals (8)

IMPACT → simulates → Pompeian Dinner

confidence 95% · We instantiate IMPACT in a five-hour simulation of a Pompeian dinner involving ten agents across interdependent roles.

IMPACT → usesmodule → Wait-or-Prompt Resolution

confidence 95% · IMPACT connects socially gated milestone planning, wait-or-prompt resolution, structured directive issuance, and directive integration.

IMPACT → usesmodule → Structured Directive Issuance

confidence 95% · IMPACT connects socially gated milestone planning, wait-or-prompt resolution, structured directive issuance, and directive integration.

IMPACT → usesmodule → Directive Integration

confidence 95% · IMPACT connects socially gated milestone planning, wait-or-prompt resolution, structured directive issuance, and directive integration.

IMPACT → usesmodule → Socially Gated Milestone Planning

confidence 95% · IMPACT connects socially gated milestone planning, wait-or-prompt resolution, structured directive issuance, and directive integration.

IMPACT → developedby → Cornell University

confidence 90% · Authors affiliated with Cornell University introduced IMPACT.

IMPACT → evaluatedby → Archaeology Experts

confidence 90% · Interviews with six archaeology experts highlighted historically plausible movement patterns...

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Simulations of archaeological sites can make interpretations of past cultural practices observable and examinable. Generative multi-agent simulations offer a bottom-up approach to modeling how people collectively moved through and used historical spaces. However, current agents designed to simulate everyday life often plan and act independently, limiting their ability to capture how movement depends on others' actions. We introduce IMPACT (Interdependent Movement Planning through Inter-Agent Constraints and Triggers), an architecture that uses culturally specific roles and obligations to define dependencies among agents' activities and guide coordination. IMPACT connects socially gated milestone planning, wait-or-prompt resolution, structured directive issuance, and directive integration. These mechanisms determine whether and when activities can begin or change as social conditions evolve, producing socially constrained and prompted movement as their primary observable outcome. We instantiate IMPACT in a five-hour simulation of a Pompeian dinner involving ten agents across interdependent roles. Analysis of five simulation runs shows how social roles, responsibilities, and status relations shape household activities and spatial practices, as reflected in patterns of co-location, asymmetric waiting, co-movement, and social directives. In a controlled ablation evaluation, thirty-seven participants rated the complete architecture's behavior as more socially coherent and believable than that of two reduced architectures. Interviews with six archaeology experts highlighted historically plausible movement patterns and the simulation's potential to support archaeological interpretation, while identifying areas requiring stronger historical grounding for future work.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

115,050 characters extracted from source content.

Expand or collapse full text

IMPACT: Modeling Socially Interdependent Movement in a Generative Multi-Agent Simulation of a Pompeian Household Tianqi Liu Affiliation: Department of Information Science, Cornell University, Ithaca, NY 14853, USA Nayoung Kim Affiliation: Department of Information Science, Cornell University, Ithaca, NY 14853, USA Julia Sebastien Affiliation: Department of Communication, Cornell University, Ithaca, NY 14853, USA Kathryn Gleason Affiliation: Department of Landscape Architecture, Cornell University, Ithaca, NY 14853, USA Caitlín Eilís Barrett Affiliation: Department of Classics, Cornell University, Ithaca, NY 14853, USA Andrea Stevenson Won Affiliation: Department of Communication, Cornell University, Ithaca, NY 14853, USA Abstract Simulations of archaeological sites can make interpretations of past cultural practices observable and examinable. Generative multi-agent simulations offer a bottom-up approach to modeling how people collectively moved through and used historical spaces. However, current agents designed to simulate everyday life often plan and act independently, limiting their ability to capture how movement depends on others’ actions. We introduce IMPACT (Interdependent Movement Planning through Inter-Agent Constraints and Triggers), an architecture that uses culturally specific roles and obligations to define dependencies among agents’ activities and guide coordination. IMPACT connects socially gated milestone planning, wait-or-prompt resolution, structured directive issuance, and directive integration. These mechanisms determine whether and when activities can begin or change as social conditions evolve, producing socially constrained and prompted movement as their primary observable outcome. We instantiate IMPACT in a five-hour simulation of a Pompeian dinner involving ten agents across interdependent roles. Analysis of five simulation runs shows how social roles, responsibilities, and status relations shape household activities and spatial practices, as reflected in patterns of co-location, asymmetric waiting, co-movement, and social directives. In a controlled ablation evaluation, thirty-seven participants rated the complete architecture’s behavior as more socially coherent and believable than that of two reduced architectures. Interviews with six archaeology experts highlighted historically plausible movement patterns and the simulation’s potential to support archaeological interpretation, while identifying areas requiring stronger historical grounding for future work. IMPACT demonstrates how agent-based modeling can operationalize culturally situated interpretations of past sociospatial practices. Keywords generative multi-agent systems ⋅· socially interdependent movement ⋅· archaeological simulation ⋅· cultural heritage ⋅· Pompeii 1 Introduction Agent-based simulation has long helped archaeologists study population movement, settlement, and human decision-making [1, 2, 3, 4]. Previous reconstructions have populated historical environments with autonomous characters, demonstrating the potential of simulation to make past settings and social activities visible for archaeological research and public engagement [5, 6]. Recent generative multi-agent systems (MASs) extend this approach by populating reconstructed environments with agents that form plans, act, and have dynamic social interactions over time [7, 8]. However, to our knowledge, no simulation has yet examined how a person’s movement is shaped by the presence, responsibilities, actions, and responses of other people in the group [9, 10]. Modeling collective use of space requires more than overlaying independent trajectories. Practices such as hospitality, domestic service, dining, and ritual involve people with different social roles. Whether, when, and with whom a person moves may therefore depend on others’ behavior, rather than be triggered simply by the passage of time or the completion of an individual task. We describe movement as socially constrained when it depends on the presence or actions of relevant others. Movement is socially prompted when an invitation or request causes another person to update their activity. Its observable consequence is that agents wait for, initiate, direct, and respond to one another as they move through space. Hierarchical societies pose a demanding setting because social position can largely shape how people conduct shared activities [10]. These dependencies are often asymmetric: different roles carry different expectations about waiting, initiating transitions, and whose request should change another person’s plan. Existing generative MASs for simulating everyday life do not typically represent these role-dependent social conditions as explicit constraints on movement or use authorized social directives to prompt another agent’s plan revision. Agents often organize their activities around individual plans, while social interactions arise through proximity, conversation, and information exchange [7, 11]. Related mobility models similarly emphasize individual activity selection, destinations, and movement trajectories [12, 13], without modeling social condition as important information to shape agents’ movement. Consequently, an agent may begin an activity before the required participants arrive, move without an expected invitation, or issue a request that does not affect the recipient’s plan. Research on multi-agent coordination addresses task allocation, role organization, communication, and collective performance [14, 15, 16, 17]. However, the lack of attention to modeling collective movement in shared space restricts the ability of simulations to represent how events depend on coordinated actions. Consider a formal dinner party. A guest should not move to the dining room simply because their individual plan indicates that dinner is scheduled. Instead, the guest should wait for the host to initiate the transition, prompting the guests to join. In this example, movement results from connected decisions: agents wait, coordinate, issue invitations or requests, and revise their plans in response to others. We introduce IMPACT (Interdependent Movement Planning through Inter-Agent Constraints and Triggers), an architecture for modeling socially constrained and prompted movement in generative MASs. IMPACT connects four mechanisms. Socially Gated Milestone Planning represents the social and spatial conditions under which activities may begin or transition. Wait-or-Prompt Resolution determines whether an agent should wait or initiate coordination when those conditions are unmet. Structured Directive Issuance communicates invitations and requests to other agents. Directive Integration incorporates accepted directives into recipients’ plans through contextually appropriate local changes. Together, these mechanisms operationalize social expectations as conditions that constrain and prompt each other’s movement. We selected a formal Pompeian dinner party as a historically grounded and demanding testbed for IMPACT. The setting concentrates interdependent roles, asymmetric authority, and transitions across shared domestic spaces. The simulation includes ten agents across multiple social roles: a host, guests, enslaved household servants, cooks, a dancer, and a gardener. Archaeology experts informed the spatial setting, agent roles, activity sequence, and role-specific expectations. Across reception, a garden visit, dinner service, entertainment, ritual, and farewell, agents coordinated emergent changes in activity and location. We evaluated whether explicitly modeling social constraints and prompts improved the coherence of collective behavior by comparing IMPACT with two reduced architectures with thirty-seven participants. We then examined patterns of co-location, waiting, collective movement, and social directives, and interviewed six archaeology experts about the simulation’s historical grounding, interpretive value, limitations, and potential applications. The resulting traces show patterns of waiting, initiating, following, and sharing space, making the consequences of modeled social roles and obligations available for examination. Archaeology experts identified the system’s potential for examining past sociospatial practices while identifying areas that require deeper historical grounding. Our contributions include: • The IMPACT architecture for socially constrained and prompted movement, which makes role-specific social conditions explicit in agent planning and enables agents to wait, initiate coordination, issue directives, and revise plans in response to others’ behavior. • A historically informed Pompeian dinner-party testbed, demonstrating how domestic labor, dining, entertainment, and ritual can be represented through role-specific dependencies and examined through patterns of co-location, waiting, collective movement, and social directives. • A multimethod evaluation that combines controlled comparisons with reduced architectures, movement and coordination analysis, and archaeology expert interviews to evaluate social coherence and assess the historical interpretive value and limits of the simulation. 2 Related Work We situate this work at the intersection of three research areas: movement and spatial practice in archaeological simulation, planning and movement in generative multi-agent simulations, and interdependent social planning and coordination. Together, these areas raise a central question: how can agents’ everyday movements reflect not only individual plans but also roles, obligations, and actions of others? 2.1 Movement and Spatial Practice in Archaeological Simulation Spatial practice is central to archaeological accounts of how past communities organized access, activity, and social interaction [18, 19]. Agent-based modeling (ABM) provides an important digital experimental method for translating hypotheses about past behavior into simulations and examining their emergent spatial outcomes [1, 20]. By combining behavioral assumptions with reconstructed environments, ABMs have been used to study settlement, migration, subsistence, and population dynamics [3, 2, 21, 22, 4, 23, 24]. Examples include models of how agricultural conditions shaped household settlement [25] and AncientS-ABM’s exploration of past settlement and social organization [19]. At smaller spatial scales, simulations have examined movement within reconstructed environments [26]. Roman street-network analyses connect spatial configuration with accessibility and circulation [9], while virtual reconstructions of Pompeii or other cultural heritage like Silves animate characters through environmental triggers, local stimuli, and personality traits [5, 6]. ABM in archaeology can model and explain human decision-making and social processes [2, 27]; visualizing and testing social or spatial hypotheses with domain experts [28]; extracting statistical patterns from simulated traces [29]; and making cultural practices accessible to non-experts and public audiences [30]. These applications also carry two linked interpretive risks. First, simulation outputs depend on incomplete archaeological evidence and on researchers’ choices about agent rules, social relationships, and parameters; risking overstating a result’s generality or robustness [31, 32]. Second, different model specifications can generate similar aggregate patterns, so a coherent simulation trace does not uniquely establish the historical mechanism that produced it [29]. These approaches largely explain movement through environmental conditions, individual preferences, or predefined behaviors, with less attention to how people coordinate their use of shared space. This gap is particularly consequential in socially stratified settings, where status and power shape behavior, access to space, and patterns of encounter [10, 33, 34, 35]. Archaeological simulation must therefore model how social relationships shaped when, with whom, and under what circumstances people moved. 2.2 Planning and Movement in Generative Multi-Agent Simulations Compared to rule-based [36, 37] and script-based agents [38], large language model (LLM) agents provide new opportunities to dynamically simulate everyday behavior through individual goals, memories, social relationships, and observation [7]. Agent frameworks like TravelAgent or Concordia simulate independent agent planning and movement based on environmental information [39, 40]. Park et al.’s generative-agent architecture combines memory retrieval, reflection, and planning to support agents that navigate shared environments, converse, and maintain daily routines [7]. Social coordination can also emerge when agents exchange invitations and independently plan to attend events. Related systems incorporate needs, emotions, and relationship closeness into daily activities and conversations [11]. Other approaches apply LLM agents directly to human mobility. LLMob generates activity trajectories from habitual patterns and personal motivations [12], while TrajLLM combines persona generation, activity selection, and destination prediction to produce realistic individual movement patterns [13]. However, everyday-life and mobility simulations generally organize behavior around individual plans, even when those plans incorporate information obtained from others. An agent may therefore enact a planned social activity even when an expected co-participant has not arrived. Coordination consequently rests on individual plans aligning and becomes fragile when social expectations are unmet. The resulting architectural gap concerns explicitly representing social preconditions and the ability to recover when those preconditions are not met. 2.3 Interdependent Coordination of Agents Multi-agent systems have long modeled interdependence through shared goals, task allocation, synchronization, and joint planning [14, 41, 42]. Organization-oriented frameworks further define agent responsibilities through roles, constraints, and interaction rules [15, 43]. Recent LLM-based systems extend these principles through role-specialized collaboration: ChatDev coordinates software-development tasks [16], while AgentVerse organizes agents to collaborate on reasoning, coding, and decision-making [17]. Other research examines how social structures shape agent behavior. SOTOPIA evaluates social reasoning in scenarios involving negotiation, relationships, and power differentials [44], while related studies investigate norm formation and social learning through repeated interactions [45, 46]. Research on authority and hierarchy further demonstrates their effects on social organization, cooperation, resource governance, and conversational compliance [47, 48, 49, 50]. 3 System Architecture Figure 1: Overview of IMPACT. Socially Gated Milestone Planning represents activities and their social and spatial execution conditions. Wait-or-Prompt Resolution determines whether an agent should wait or initiate coordination when a condition remains unmet. Structured Directive Issuance communicates invitations or requests, and Directive Integration incorporates accepted directives into the recipient’s plan. The revised plan re-enters milestone execution, while the recipient’s response or action may resolve the sender’s original dependency. To visualize how people with different social roles use the space together, we introduce IMPACT (Interdependent Movement Planning through Inter-Agent Constraints and Triggers), an architecture that models movement as the outcome of socially interdependent planning. IMPACT operationalizes interdependence through two complementary mechanisms. Social constraints determine whether an existing activity can proceed; social prompts seek to influence another agent’s behavior through social directives. Figure 1 shows how the following four interconnected modules implement this coordination loop. Our current implementation of all LLM-powered agents utilized GPT-5.4 [51]. 3.1 Socially Gated Milestone Planning This module translates an agent’s high-level timeline into an ordered, socially situated plan. Each milestone represents an activity together with the social and spatial conditions that must hold for the agent to proceed. The design draws on accounts of shared intention and joint action, in which individuals remain responsive to others’ actions and commitments [52, 53, 54]. Unlike a fixed-duration schedule, a milestone advances only when its observable success condition is satisfied, allowing its actual duration to vary with other agents’ behavior. Milestone construction. For agent a, a high-level, natural-language-defined timeline T is converted into an ordered milestone plan Π=[m1,m2,…,mn] =[m_1,m_2,…,m_n]. Each milestone is represented as mi=⟨gi,δ^i,i,ϕi,i⟩,m_i= g_i, δ_i,C_i, _i,A_i , (1) where gig_i is its objective, δ^i δ_i its estimated duration, iC_i its socio-spatial configuration, ϕi _i its observable success condition, and i=[α1,…,αk]A_i=[ _1,…, _k] its sequence of finer-grained actions, each spanning approximately three to five minutes. The configuration i=⟨zi,i⟩C_i= z_i,P_i specifies the spatial region ziz_i and required participants iP_i. A new activity milestone is introduced when either component changes. For example, waiting for guests and greeting them after they arrive are separate milestones because the required participants change, even though the region remains the same. Greeting guests at the entrance and leading them to the reception room are also separate milestones because the region changes. Actions that preserve both components remain within the same milestone. Navigation milestones. When consecutive activity milestones involve different regions, a dedicated navigation milestone is inserted between them. Navigation begins after the preceding activity succeeds and ends upon physical arrival, expressed as ϕnav(ot)=[z(ot)=zj] _nav(o_t)=I[z(o_t)=z_j], where oto_t is the current observation, z⁡(ot)z(o_t) the agent’s current region, and zjz_j the destination. If travel takes longer than estimated, navigation continues until arrival, preventing the destination activity from beginning prematurely. Transition success conditions. For each adjacent pair of activity milestones (mi,mi+1)(m_i,m_i+1), the system infers the observable condition ϕi _i that completes mim_i and activates mi+1m_i+1. The LLM examines milestones’ objectives along with the sociospatial configuration change from iC_i to i+1C_i+1. When the region remains the same, but the required participants change (zi=zi+1z_i=z_i+1 and i≠i+1P_i _i+1), ϕi _i typically concerns the arrival or departure of a relevant agent. When the next activity takes place in a different region (zi≠zi+1z_i≠ z_i+1), ϕi _i specifies what makes departure from the current region appropriate, such as issuing or receiving a moving invitation, or notifying a relevant agent of an intended departure. If the transition introduces no social dependency, ϕi _i may be satisfied by reaching the estimated endpoint of the current activity. When all plans in iA_i are completed while ϕi _i remain unmet, the agent enters the wait-or-prompt transition-resolution process. 3.2 Wait-or-Prompt Resolution When the estimated duration of milestone mim_i has elapsed while its success condition ϕi _i remains unsatisfied, this module determines whether agent a should continue waiting or take a coordination action intended to help satisfy ϕi _i. The decision considers the current and next milestones, the current observation oto_t, prior executed action iA_i, the prior resolution decisions of the current milestone t,iW_t,i, and the agent’s role-and-responsibility profile ρa _a. Two types of dependencies enable active coordination. A co-present movement dependency arises when relevant agents are already in the current region and must coordinate a transition. A missing-participant dependency arises when satisfying ϕi _i requires an agent who is not currently present. Other unmet conditions, such as awaiting an invitation or another agent’s departure, result in continued waiting without invoking a wait-or-prompt decision in the current implementation. For co-present movement dependency, IMPACT selects among InviteCoMove, InformMoving and Wait. InviteCoMove asks relevant agents to accompany the initiator to the next region. In contrast, InformMoving notifies relevant agents of the initiator’s intended departure without asking them to change their own plans. Thus, InviteCoMove may generate a coordination intention for directive issuance, whereas InformMoving remains a non-directive notification. For a missing-participant dependency, IMPACT selects among DelegateFindAndDirect, FindAndDirectPersonally and Wait. DelegateFindAndDirect directs another agent to locate the missing participant and communicate the directive. FindAndDirectPersonally adds a search action to the initiator’s plan, enabling them to locate and direct the missing participant. In either case, Wait extends the current milestone and generates additional actions consistent with its existing objective. A non-wait response instead augments the current action sequence with resolution-oriented actions. If ϕi _i becomes satisfied during this extended execution, the agent advances; otherwise, the module reevaluates the decision when the extended duration elapses. Any response that requires another agent to revise its plan forwards a coordination intention to Structured Directive Issuance. 3.3 Structured Directive Issuance When a coordination response may need another agent to revise its plan, this module determines whether the acting agent is authorized to express that need as a directive and, if so, encodes it in a structured form. The design draws on deontic accounts of social interaction, which explain how an utterance can establish what another participant is expected or entitled to do next [55]. Before initiating a conversation, agent a evaluates whether completing its current action αj∈i _j _i requires another agent to change their plan, and whether its role-and-responsibility profile ρa _a authorizes it to seek that change. If no authorized coordination intention is identified, the language-model prompt is constrained to generate dialogue without a social directive to prevent unintended plan changes. When an authorized coordination intention is identified, the system constructs a directive with information that the recipient needs to interpret and/or integrate the proposed plan change: d=⟨s,ℛ,τ,z,α⟩,d= s,R,τ,z,α , (2) where s is the sender, ℛR the intended recipients, τ the directive type, z the target region, and α the requested action or objective. Following accounts that classify requests and invitations as directive actions [56], IMPACT distinguishes τ∈,τ∈\ Req, Inv\. A Req asks a recipient to carry out an instrumental action, such as locating another agent. In contrast, an Inv asks one or more recipients to participate in a shared activity or coordinated movement. For example, asking a servant to locate the dancer is a Req, whereas asking arriving guests to accompany a servant to the reception room is an Inv. The system realizes d as an utterance that communicates the requested action, relevant destination, and social purpose. It then passes the information in conversation to Directive Integration, where the recipient may seek clarification or accept the directive, and change plan accordingly. 3.4 Directive Integration From the directive recipients’ perspective, dialogue establishes shared interpretations, and accepted directives are incorporated into their plans. Directive identification and acceptance. For a potential recipient r, let Π denote its milestone plan, oto_t its observation, HtH_t its current relevant memory to the sender, ⁡(ot)V(o_t) the agents currently visible to it, and Z the set of valid regions. During dialogue, a recipient who recognizes a possible directive requests clarification if the sender, intended recipient, destination, or requested action is unclear; otherwise, it explicitly accepts or declines the directive. After the interaction concludes, accepted directives are identified using IdentifyDirective(Ht,r,⁡(ot),) IdentifyDirective(H_t,r,V(o_t),Z). A directive applies to r only if r∈ℛr ; ambiguous, declined, or misaddressed directives do not modify its plan. The interaction is recorded in both agents’ histories, allowing responses to satisfy the sender’s original success condition. Plan revision and routing. For each accepted directive, the system selects the smallest plan update needed to represent the resulting obligation. ⁡(mj) ActivateExisting(m_j) activates an existing milestone that already covers the requested objective, preventing duplicate commitments. ⁡(α) InsertAction(α) adds the requested action to the current milestone when its region and participant configuration remain appropriate. ⁡(mnew,λ) InsertMilestone(m_new,λ) inserts and activates a new milestone when the directive requires a different spatial region or participant configuration. Any required movement is handled through the navigation milestones defined above. Additional operational details and a fully specified directive example are provided in Appendix A. 4 An Experimental Study We instantiated IMPACT in a Pompeian household scenario with ten agents to examine how historically situated social roles shape interdependent plans and movement. We evaluated IMPACT through behavioral analyses across repeated simulation runs, social plausibility compared to controlled system ablations, a module-level agreement rating, and interviews with archaeology experts [57, 58, 59]. We address three research questions: • RQ1: Can IMPACT realize a complex Pompeian household practice—here, a dinner party—through socially believable collective agent activity and movement, and how do human participants evaluate this activity and movement relative to that produced by reduced architectures? • RQ2: What interdependent activity and movement patterns emerge from agents’ culturally situated roles and obligations, and how do these patterns reflect the spatial organization of the modeled Pompeian household practices? • RQ3: How do archaeological researchers assess the historical plausibility and interpretive value of the case study and generated movement patterns, and what limitations or directions for further development do they identify? 4.1 Historical and Spatial Setting Based on the expertise of the fourth and fifth authors, archaeology experts on CRC, we simulated a five-hour dinner-party scenario set in the Casa della Regina Carolina (CRC), a Roman household in Pompeii. The environment was reconstructed as a two-dimensional pixel map based on the archaeological floor plan (Figure 2). We translated the floor plan into a navigable spatial layout using Tiled [60], identifying the functions, accessibility, and historically plausible arrangement of the household’s major areas. Using historical visual references, we generated pixel-style furnishings and textures with Gemini Nano Banana 2 [61], and assembled them into spatial regions with distinct functions, including the reception room, where the guests wait for the host and the host greets arriving guests; the garden; the triclinium, or dining room; the tablinum, or host’s office; the household shrine; the kitchen; the servants’ quarters; and other household work areas [62, 63]. Agents navigated between these regions but could perceive only the objects, people, and activities within their current region. Information about events elsewhere in the household therefore had to be acquired through movement or from other agents. Figure 2: Reconstruction of the Casa della Regina Carolina simulation environment. Spatial region functions informed by archaeologists. 4.1.1 Dinner-Party Scenario We selected a formal dinner party for two reasons. First, dinner parties involve multiple social groups, activities across different household spaces, and close coordination among hosts, guests, and household workers. Second, previous scholarship provides a basis for participants’ roles and obligations, enabling assessment of the simulation’s historical plausibility. The simulated dinner party progressed through six major phases: guest reception, a garden visit, dinner service, entertainment, household ritual, and farewell. Each phase required agents to form appropriate social configurations in particular locations. Guests gathered in the reception room, accompanied the host to the garden, assembled in the dining room for dinner and entertainment, and then proceeded to the household shrine. Meanwhile, household workers coordinated complementary activities: receiving guests, preparing and serving dinner, and summoning the dancer. For the complete simulation context and phase-level timeline, see Appendix B.1. 4.1.2 Agents, Roles, and Information Asymmetries The scenario involved ten generative agents: one host (Dominus), one honored guest (Amicus), one client guest (Client), three enslaved household servants (Servus), two cooks, one gardener, and one dancer. We assigned agents to four status levels. We classified the cooks, gardener, and dancer as specialized household workers at level 1; the three household servants at level 2; the client guest at level 3; and the host and honored guest at level 4. We incorporated these status levels into each agent’s role-and-responsibility profile ρ and used them as contextual information in prompts for role-sensitive decisions. Each agent received a role-specific activity description defining its initial knowledge, responsibilities, and anticipated activities. Knowledge of the dinner party was distributed asymmetrically and without specifying how agents should coordinate. The host knew the overall sequence of events; the guests were not informed of subsequent activities; and household workers received only the information relevant to their respective responsibilities. For example, guests could not independently anticipate when to leave for the next region, and a worker responsible for summoning the dancer could act only after receiving or otherwise acquiring the relevant information. Thus, agents needed to observe one another, wait for required social conditions, communicate invitations or requests, and revise their local plans in response to others. The complete agent-specific activity descriptions are provided in Appendix B.2. In addition to IMPACT modules, we implemented other functional modules that enable agents to perceive, store memory in the database, retrieve relevant memory from keywords, and navigate using an A* algorithm, following the work of Smallville [7]. The GitHub repository for our project is available at: upon paper publication. 4.2 Participant Evaluation We conducted a two-part participant study to answer RQ1, evaluating the perceived social believability of the simulated behavior and movement. First, we compared the observable behavior produced by Full IMPACT with that of two reduced but functional control architectures. Second, we examined whether LLM-selected decision points in the four IMPACT modules were perceived as reasonable given each agent’s role and context. We report the condition comparison below and present the module-level evaluation and its results in Appendix D. 4.2.1 Participants We recruited 37 non-expert participants through email and flyers. They were paid $10.00 USD for the study and signed the consent form approved by our institution’s IRB. Participants were 19–47 years old (M=27.16M=27.16, SD=6.30SD=6.30). Eighteen identified as female (48.6%), 17 as male (45.9%), and 2 preferred not to disclose gender (5.4%). Two participants self-identified as Bi-racial, Mixed, Multicultural, or Multi-racial, one as Black American, five as Caucasian or White/of European ancestry, 19 as East Asian, one as Hispanic, Latino, Chicano, or Puerto Rican; five as South Asian; four as Southeast Asian and one self-described as “Mediterranean".Regarding their prior knowledge of Roman history and society, 10 participants (27.0%) reported little or no knowledge, 16 (43.2%) reported basic knowledge, and 11 (29.7%) reported moderate knowledge. Figure 3: Mean participant ratings of Full Architecture, Control 1, and Control 2 across five measures. Ratings ranged from 1 to 7, with higher scores indicating more favorable evaluations. Error bars represent 95% confidence intervals. 4.2.2 Conditions We created two functional control conditions through module ablations. All configurations used the same environment, characters, initial locations, context, and language model. Full IMPACT. All four modules were enabled. Control 1: No Milestone. We disabled socially gated milestone planning. Agents progressed by elapsed time only. To prevent missing information from mechanically blocking the simulation, guest agents received the same complete timeline information as the host. Control 2: No Wait-or-Prompt. Milestones remained enabled, but agents could not select Wait when social preconditions remained unmet; they were instead required to initiate coordination. Disabling Directive Issuance, Directive Integration, or agents’ ability to select Prompt prevented the overarching event sequence from completing. Because these variants did not produce complete, comparable trajectories, we excluded them from the participant evaluation. We generated five complete Full IMPACT runs and two complete runs for each control, yielding nine complete simulation runs in total. 4.2.3 Procedure The study followed a condition-blinded, within-subjects design. Participants selected three events from a chronological list of ten that collectively captured major social activities across the dinner party. Appendix C.1 provides the complete list. For each selected event, participants received a brief description of the social situation and focal agents. Participants then viewed three clips for each selected event, one from each condition. The Full IMPACT clip was randomly sampled from the five complete Full IMPACT runs, whereas each control clip was randomly sampled from the two complete runs. We randomized the clip order within each event for each participant. For each clip, we provided suggested timestamp ranges and navigation controls so participants could find relevant interactions without watching the full five-hour simulation. 4.2.4 Measures and Analysis After viewing each clip, participants rated the focal agents on five dimensions: movement rationale, transition timing, bottleneck resolution, group synchronization, and overall social believability. All items used seven-point Likert scales from 1 (Strongly disagree) to 7 (Strongly agree). Complete instructions, event descriptions, item wording, and statistical results are consolidated in Appendix C.2 to C.6. Because participants evaluated all three conditions but only a subset of events, the study followed a within-subjects design with incomplete event blocks. Responses marked Not applicable were excluded from the corresponding dimension-specific analysis. For each dimension, we fitted a linear mixed-effects model with condition and event as fixed effects and participant as a random intercept. Full IMPACT was treatment-coded as the reference condition. Model-based pairwise comparisons were Holm-adjusted within each dimension. 4.2.5 Results Full IMPACT received the highest ratings on all five dimensions (Figure 3). Holm-adjusted comparisons confirmed that it received significantly higher ratings than both controls on every dimension. Control 2 also received significantly higher ratings than Control 1 across all five dimensions (all p≤.0164p≤.0164). The largest estimated reduction for Control 1 occurred in bottleneck resolution, followed by group synchronization. This pattern suggests that time-driven execution without socially gated planning was particularly ineffective when agents needed to coordinate co-movement. Although Control 2 retained prompting mechanisms to resolve social dependencies, its ratings remained significantly below those of Full IMPACT. Requiring agents to initiate coordination may have reduced social plausibility when their roles or the social context called for continued waiting. Together, these results indicate that socially believable movement benefited from the complete coordination loop and answer RQ1 by showing that participants perceived the full architecture as producing the most coherent, synchronized, and socially believable behavior. 4.2.6 Case Study: Directive-Mediated Coordination Figure 4: A coordination trace showing how a blocked social milestone is resolved through wait-or-prompt resolution, directive issuance, and recipient-side plan revision, ultimately making the planned activity executable. Figure 4 illustrates how IMPACT’s four modules jointly resolved a stalled entertainment event. The Host’s planned transition from “having dinner with guests" to “enjoying the dance with guests" remained blocked because the Dancer, whose presence was required, was waiting in another region. Through Wait-or-Prompt Resolution, the Host selected Prompt rather than Wait and issued a directive asking an available Servant to summon her. The Servant grounded this request in his own plan, located the Dancer, and issued a second directive conveying the Host’s request. The Dancer then accepted and grounded the directive. Because performing was already represented as a future milestone in her plan, the system activated that milestone without creating a duplicate. She moved to the Triclinium and performed, satisfying the social condition for the Host’s transition and allowing the event to proceed. The same process could also introduce an unanticipated obligation. In another trace, a directive asked the Gardener to distribute wreaths, an activity absent from his original plan. The gardener inserted a new wreath-giving milestone into his existing milestone sequence and moved to the required location to perform it. Together, these cases demonstrate that directive integration can either activate an anticipated obligation or incorporate a new one through localized plan revision. 4.3 Emergent Interdependent Movement Patterns Figure 5: Patterns of socially constrained and prompted movement generated by IMPACT. Across five independent five-hour simulation runs, A. Co-Location shows the mean proportion of time that each agent pair occupied the same region; B. Waiting Dependencies show the cumulative duration for which each row agent waited for the corresponding column agent; C. Co-Movement shows the number of events in which each row agent initiated movement and the column agent followed; and D. Social Directives show invitations (brown) and requests (blue) issued by each row agent to each column recipient, classified from the recipient-side records. Bubble size and color intensity encode magnitude, and labels report the corresponding percentage, duration (minutes), or event count. Empty cells indicate that no corresponding relationship was observed. In this section, we examine four complementary patterns to answer RQ2: What patterns of movement emerge from culturally situated roles and obligations, and how do these patterns spatially reflect the Pompeian household practices? Figure 5, panel A shows the spatial co-location configurations; Panel B identifies whose actions constrain others’ transitions; and Panels C–D show how agents prompt movement through co-movement initiation and social directives. Together, these analyses reveal how collective space use emerges from inter-agent dependencies rather than from the aggregation of independently executed trajectories. Co-Location Panel A reports the proportion of the dinner event during which each dyad occupied the same region, averaged across five simulation runs. Co-location closely followed agents’ social relationships and functional roles. The host shared a spatial region with each guest for 82.9% of the event, while the two guests were co-present for 93.3%, indicating that the principal party moved through the house as a coordinated social unit rather than as independent agents. The two cooks remained together throughout the simulation (100%), reflecting their shared workplace and closely coupled tasks. By contrast, co-location among the three servants was uneven: Servants 1 and 3 were together for 63.5% of the event, whereas the other servant dyads were together for only 11.2–14.4%. Thus, agents with similar status did not necessarily follow identical trajectories; different responsibilities distributed them across spaces. Spatial separation between elite agents and lower-status labor was also selective. The host and guests frequently shared space with the servants who directly supported the dinner (45.6–52.2%), but had no observed co-location with the cooks and relatively limited co-location with the dancer and gardener (16.2–26.3%). This pattern distinguishes visible, guest-facing service from less visible back-of-house and peripheral labor, making the modeled organization of domestic work spatially legible. Waiting Dependencies Panel B shows the cumulative time for which agent i waited for agent j across the five simulation runs. Waiting was strongly asymmetric. The guests had no recorded waiting dependencies, and the host waited only for Servant 1 to announce the guests’ arrival (60 minutes cumulatively across the five runs). Most waiting was instead concentrated among servants and the dancer, whose planned transitions depended on the arrival, readiness, or actions of higher-status or coordinating agents. IMPACT therefore makes social constraints observable as a temporal dependency: an agent’s next activity does not begin simply because its individual schedule reaches that point, but remains gated until another agent establishes the required social condition. Co-Movement Panel C reports how often agent i initiated and led a coordinated movement that agent j subsequently followed, aggregated across five runs. Co-movement was most frequent within the host–guest group. The host led each guest on 19 occasions, confirming that these three agents formed the most consistently synchronized moving group in the dinner event. The host initiated the largest number (45) of co-movement events overall, and 38 of these involved the guests. Servant 1 initiated fewer events but coordinated the widest range of agents, leading guests, other servants, the dancer, and the gardener. Co-movement thus also makes social constraints observable by revealing two forms of initiative: the host organized the movement of the principal social party, while Servant 1 coordinated between groups. Social Directives Panel D shows the social directives issued by agent i and grounded by recipient j, distinguishing invitations from requests and aggregating counts across five runs. Because directive type was extracted from the recipient’s interpretation, these categories represent how directives were grounded by receivers and may not exactly match an issuer-side classification. When receiving a directive from a higher-status agent, recipients more often recognized it as a request than as an invitation. The host issued the largest number of directives, followed by Servant 1, again identifying them as the simulation’s two principal coordination hubs. Their roles, however, differed. The host rarely communicated directly with cooks, the dancer, or the gardener. Instead, Servant 1 issued directives across servants and lower-status workers, thereby relaying coordination to agents outside the host’s immediate interaction network. The resulting structure was therefore mediated rather than simply top-down: the host established high-level social transitions, while servants—especially Servant 1—propagated the requests and invitations required to implement them. Taken together, the four panels reveal a coherent organization of interdependent movement. Co-location shows the resulting social configuration of space; waiting identifies who is constrained by others; and co-movement initiation and directives identify who prompts transitions and how those prompts propagate across agents. IMPACT thus allows movement to emerge as a collectively accomplished process shaped by social dependencies, differentiated responsibilities, and cross-agent plan revision. 4.4 Expert Evaluation 4.4.1 Participants We recruited six archaeology experts, defined as individuals who had completed or were currently enrolled in a graduate-level academic qualification in archaeology, classics, landscape architecture, or a closely related field focused on ancient Roman history, material culture, or built environments. Participants were 22–69 years old (M=34.83M=34.83, SD=18.82SD=18.82); five identified as female and one as male. All six participants identified as White or of European ancestry, and one participant additionally identified as East Asian. Four participants were graduate students, one was a retired professor, and one was a full-time professor. Their primary research areas spanned Roman gardens, early Imperial archaeology and architecture, Latin epigraphy, domestic and household archaeology, Pompeii, religion and ritual, and ancient Mediterranean material culture. Four participants reported advanced knowledge of Roman history and society, and two reported expert knowledge. 4.4.2 Procedure Each semi-structured interview lasted approximately 60 minutes. Interviews were transcribed using Zoom’s automated transcription software. Experts viewed the complete simulation randomly picked from the five runs and were asked to discuss: 1. the social and historical plausibility of the agents’ behavior, dialogue, and use of space; 2. potential applications and intended audiences for the simulation (i.e., for archaeological research, education, heritage interpretation, and public outreach); and 3. risks and recommendations surrounding presenting a generative simulation of the past. The semi-structured interview guide was attached in Appendix E. Experts then evaluated events from the ablated conditions using the same rating system as novice participants while thinking aloud. P2–P4 each reviewed three self-selected events; P1, P5, and P6 each reviewed one due to time constraints. 4.4.3 Analysis The second and third authors coded all expert interviews in Atlas.ti [64]. A hybrid deductive-inductive coding approach was employed. Eleven a priori codes were developed based on the research questions: social plausibility, social implausibility, historical plausibility, historical implausibility, recommended uses, non-recommended uses, recommendations to improve social plausibility, recommendations to improve historical plausibility, power/status/hierarchy, positive affect, and negative affect. During analysis, the authors met to discuss emerging themes and refine the coding framework. Three additional codes emerged: technical features/user interface, expert uncertainty, and potential harms/risks. The transcripts were subsequently revisited to ensure these codes were applied consistently across the dataset. 4.4.4 Findings Of the fourteen themes identified through coding, most converged around three broader themes: (1) historically and socially plausible social and spatial patterns; (2) limitations to the historical and social plausibility of behaviors, including potential harms and risks; and (3) recommended applications, use cases, audiences, and future directions. Co-Location Experts identified status-based spatial segregation as historically plausible. For example, P6 appreciated that the cooks were less visible, as this “correspond[s] to the Roman idea that cooking itself was unsightly, and that the people who did it were smelly because of their proximity to the food and to the latrine." Further emphasizing the historical importance of hierarchical non-co-location, P1 recommended “exaggerat[ing] the social distance" between individuals of different statuses. Specifically, P1 explained that “once the guest arrives, enslaved people [should] start to become somewhat invisible."To this end, P1 suggested limiting direct service within the triclinium to a single higher-status servant and having lower-status servants serve from beyond the surrounding walls which, in reality, were only knee-high, thereby “giv[ing] a little more space around the elite [characters]." P1 additionally argued that social and spatial differentiation should include status hierarchies among servants and enslaved people themselves. As P1 explained, “the people who have worked longer and are older … simply will not do certain jobs; they will make the younger ones do it." Waiting Dependencies Waiting behaviors generally enhanced the plausibility of the simulation. Most obviously, waiting for relevant events to occur was often necessary for a scene to make logical sense; otherwise, there could occur a “hallucination-like moment [such as] when no performer was actually there, but everyone was enjoying and commenting on the performance," in an ablation condition (P4). Experts particularly viewed waiting favorably when it reflected Roman social norms and status hierarchies. For example, P4 considered aspects of an ablation condition “reasonable" so long as servants waited for instructions from the dominus (P4). Indeed, waiting for the dominus to act first was so central to P4’s evaluation that it formed the basis of their assessment of the historical implausibility of one ablation condition in which a guest “wander[ed] around the host’s house" (P4). As P4 explained, “A guest should normally wait for the host’s invitation or guidance before moving to another part of the house. Otherwise, leaving on their own could be impolite" (P4). Similarly, P2 appreciated the simulation’s depiction of the host remaining in a private space while enslaved servants received and managed arriving guests, which P5 called “very realistic" and “very consistent with the historical context" (P5); as P2 explained, “it’s really accurate, the fact that the master of the house didn’t greet immediately the host at the entrance of the house, but waited in his own room and let the slaves take care of it. … if you were rich enough, you could have a slave whose only job was to do that" (P2). However, experts also expressed uncertainty regarding whether particular waiting behaviors accurately reflected ancient Roman practice. For example, the simulation depicts the dancer waiting in a designated servant’s preparation room until summoned by the dominus. Several experts viewed this behavior as historically plausible, particularly in comparison to ablation conditions in which the dancer failed to wait for a cue by an emissary of the dominus (P3, P4, P6). P1, however, disagreed, noting that accomplished performers may instead have entered alongside musicians with theatrical flourish, accompanied other participants to the shrine, or otherwise remained visible prior to their performance. By contrast, lower-status performers may simply have waited in place prior to the guests’ arrival, rather than remaining in a separate preparation room until summoned, as depicted in the simulation. Nevertheless, P1 argued that, regardless of its historical accuracy, this waiting mechanic served an interpretive purpose by drawing attention to the existence of another room within the house and signaling that the dancer’s work had concluded when they returned there. At other times, however, waiting behaviors appeared to diminish both historical plausibility of the simulation by reflecting contemporary assumptions rather than ancient Roman practices. As P6 pointed out, there “might have been a miscommunication about what a ’reception room’ is for … in Roman society"; unlike its contemporary meaning as a space in which guests await the host, “when we archaeologists talk about rooms as for reception, that doesn’t mean waiting until the host gets there; it means a place where the host would hang out with people and where socializing would happen." Co-Movement Co-movement often contributed to experts’ views of the simulation’s historical plausibility. For example, P4 said: “The host invited the guests, they accepted, and everyone went together. I thought that was very reasonable." Often, co-movement patterns were viewed as particularly historically valuable when they were shaped by implicit social directives derived from status hierarchies. For example, when the host initiated a movement and the servant followed him, that seemed normal to P4, even if the host had not explicitly told the servant to follow him: “By comparison, it would feel inappropriate for the servant to hurry the host along or simply leave on his own" (P4). Social Directives Experts frequently viewed explicit social directives as historically plausible when they reflected Roman status hierarchies. In the simulation, the dominus issued directives either directly or indirectly through servants, a pattern that several experts regarded as historically appropriate (e.g., P1, P3, P4). As P5 said approvingly, “I could … tell, for example, that the host had a higher social status, while the servants were more respectful and followed the host’s instructions." Accordingly, ablation conditions in which directives violated these hierarchies were often perceived as historically implausible. For example, P4 critiqued an ablation since, “The servant should not have acted on his own before the host had even spoken." P4 further commented on another ablation condition: “in this [ablation] scene, the servant told the host and guests to go to the Shrine and wait there. My reaction was almost, ’How dare you? How could you overstep your position and status like that?"Similarly, P2 criticized an ablation condition in which enslaved servants abruptly instructed everyone to leave, commenting that “it’s more believable if the two people with the higher status took that decision by themselves" (P2). Recommended Applications Other findings clustered around recommended applications, use cases, and audiences for the simulation. Experts generally discussed its value as a tool for communicating knowledge about ancient Roman society, material culture and site-specific finds; however, P6 also mentioned a second desirable goal for the simulation in parallel: the generation of new knowledge and insights into ancient Roman life for experts. Application 1: Communicating Archaeological Knowledge Experts most often discussed the ease and utility of using the simulation as a communicative tool “for engaging with a broader public and communicating things we already know" about ancient Roman society and material culture (P6). Three primary audiences emerged from these discussions, distinguished largely by their level of expertise. Members of the general public (P1, P5, P6) could be reached through museum exhibitions and on-site installations presented alongside archaeological remains (P2, P4). Several experts (e.g., P3, P5) suggested that the simulation’s relatively limited and “surface-level" interactions made it an effective way “to introduce the archaeology world" to “a broader audience" (P3, also P5) while serving “a pedagogical function in showing people more about how Roman social relations worked" (P6). P5 similarly emphasized the simulation’s ability to communicate the ancient world in a “vivid, concrete, and visual way" that could “leave a stronger impression on [novice] visitors" (P5). For students at or below the college level (P1, P2, P3), experts suggested that the simulation could support learning among archaeology majors who had not yet developed expert knowledge and could help prepare students for participation in site excavations (P2). P1 further identified experts as a potential audience, suggesting that the simulation could be used to communicate site-specific archaeological findings to professional communities.For these audiences, limited character interactions and minor inaccuracies were not seen as especially problematic, as they would likely be recognized immediately and could even help draw attention to particular site features and findings (P1). Application 2: Generating Novel Insights into Ancient Roman Society and Material Culture P6 suggested that, because the simulation continuously models agents’ movements and activities over time, it may be less dependent on researchers’ selective attention and therefore can generate new insights into overlooked aspects of everyday life. For example, observing that agents typically took the most direct route to their destinations prompted P6 to consider how patterns of movement through space may have shaped social interaction when practical movements through space temporarily disrupted Roman ideals of social segregation. P6 reflected on how such moments of co-location may have given rise to social interactions that are largely absent from both ancient sources and modern scholarship: “What if one of the guests had to go to the bathroom? They would have basically had to pass through the kitchen, and in the process, they would have seen the cooks at work. What sort of social interactions, if any, does that facilitate?" Risks and Recommendations Despite the simulation’s perceived value, experts identified several potential risks associated with the use of generative multi-agent simulations of cultural heritage sites and social dynamics and proposed recommendations for mitigating them. Several experts remarked that the simulation exhibited only a limited set of activities, lacking the richness, spontaneity, depth, and variability characteristic of everyday interactions with one’s social and material surroundings (P4, P5). To mitigate the risk of oversimplifying the complexity and dynamism of the ancient Roman world, experts suggested incorporating greater opportunities for user agency, allowing users to introduce their own actions, questions, and lines of inquiry and thereby fostering more varied and dynamic interactions in a role-playing experience (P1, P4, P5, P6). Experts repeatedly expressed concern that generative agent systems may reproduce cultural biases in their representations of ancient Rome stemming from their training data. For example, experts remarked that the dialogue sounded more Victorian or Shakespearean than Roman in tone (P1, P6). P6 argued that such associations are not politically neutral, explaining that popular depictions of Romans speaking with British accents stem in part from efforts to portray the British Empire as the heir to Rome. As P6 noted, “if we buy into that, then we are reproducing Victorian propaganda, basically," emphasizing the importance of challenging such assumptions (P6). To this end, most experts recommended grounding simulated dialogue and behavior more directly in historically informed source materials (P1, P3, P5, P6), including both secondary scholarship on ancient Rome (P3, P5) and primary sources, i.e., high-quality translations of literary works by contemporary authors (P1, P6). As P6 qualified, “that wouldn’t necessarily tell us what actual Romans would have done, … [as] it would be sort of equivalent to an AI generating new scripts for Seinfeld"; nevertheless, they could still produce behaviors and conversations that were broadly “consistent" with, and “plausible" for, the historical context (P6). Others proposed creating dynamic, source-based scenes tied to specific historical texts (P1). By contrast, P6 advocated either embracing historical authenticity through the use of untranslated Latin or abandoning attempts to reconstruct a specifically Roman voice altogether in favor of a deliberately modern and accessible colloquial style. Several experts also expressed concerns that novice and non-expert audiences might mistake plausible AI-generated reconstructions for established historical fact. The simulation might portray events unsupported by available evidence; or, due to uncertainties and competing interpretations of the historical record, the simulation must often extrapolate beyond what is currently known and generate one plausible reconstruction among many. To mitigate these risks, participants proposed a range of back-end strategies, including incorporating stronger fact-checking and verification procedures to ensure that simulated events align with available evidence (P5). Other participants proposed audience-facing solutions, such as signposting the known and unknown functions of objects and rooms and their sources using informational overlays or hover-text descriptions (P4). Others argued that simulations should explicitly avoid portraying generated content as historical fact altogether (P3): “It is generated. It is close to the facts. But after looking at the visualization, people may easily believe…’This is how it was in the past.’ 5 Discussion This study introduced IMPACT, an architecture that models individual movement as an outcome of socially interdependent activity. Our results demonstrated how coordination takes different forms across roles and status. Our findings from ablation conditions suggest that agents’ social behavior in IMPACT, specifically regarding (A) Co-Location, (B) Waiting Dependencies, (C) Co-Movement, and (D) Social Directives, was socially plausible; meanwhile, expert interviews offered supporting confirmatory evidence that this social plausibility can provide a foundation for historical plausibility in an ancient Roman context, and suggested applications for both communicating and generating knowledge about the ancient world. This overlap suggests that coherent social coordination is an important basis for historically plausible behavior. Historical plausibility, however, places additional demands on how agents interact. Such expectations go beyond general social coherence and require knowledge of specific historical practices. Domain expertise is therefore needed both to evaluate simulated behavior and to ground agent roles, interactions, and movement in the historical context [58]. 5.1 Limitations and future directions This study applied IMPACT to one expert-informed reconstruction of a Pompeian household dinner, with simplified representations of roles, status, relationships, and knowledge distribution. The findings are therefore specific to this historical interpretation and the scale of the scenario. In addition, biases in language models’ training data and developers’ knowledge and modeling choices may introduce modern cultural assumptions into depictions of the ancient world, particularly when historical evidence is limited [65]. A simulation can therefore be socially coherent but historically implausible. Expert input could strengthen the links between historical evidence and model behavior. Future work could link specific modeling choices to textual sources, archaeological findings, or scholarly debates, then use sensitivity analyses to examine how changes to those assumptions affect the results [31]. Curated historical materials and role-specific examples could also inform agent dialogue, with expert evaluation used to assess whether they improve its historical plausibility. Comparisons across sites, social settings, and competing historical interpretations could help distinguish patterns that depend on particular assumptions from those that recur across contexts. Future studies could examine larger groups, changing roles, longer time periods, and activities spanning multiple households or public spaces. Models could also more finely distinguish among enslaved workers, performers, and guests, including their responsibilities, freedom of movement, and access to information. Researchers could vary these factors and then examine changes in co-location, waiting, and movement. These comparisons would make the spatial consequences of different historical assumptions visible and support their systematic interpretation and analysis [27]. 6 Conclusion We introduced IMPACT, a generative multi-agent architecture that makes movement socially constrained and responsive to other agents. Its mechanisms produced differentiated patterns of co-location, waiting, co-movement, and directive exchange, illustrating how social status and responsibilities can shape household activities and spatial use. Archaeological researchers further highlighted the interpretive value of these interdependent movement patterns, while identifying risks and priorities for deeper historical grounding. 7 Acknowledgments Tianqi Liu and Julia Sebastien thank the Casa della Regina Carolina (CRC) project for the opportunity to participate in the excavation in Pompeii and the archaeologists there for their insights. ChatGPT (GPT-5.6 Sol, OpenAI) was used for grammar checking and language editing on the final draft. All AI comments were carefully reviewed by the authors. References [1] S. Graham (2022) An enchantment of digital archaeology: raising the dead with agent-based models, archaeogaming and artificial intelligence. Berghahn Books. Cited by: §1, §2.1. [2] J. A. Barceló and F. Del Castillo (2016) Simulating the past for understanding the present. a critical review. Simulating prehistoric and ancient worlds, p. 1–140. Cited by: §1, §2.1, §2.1. [3] W. H. Cegielski and J. D. Rogers (2016) Rethinking the role of agent-based modeling in archaeology. Journal of anthropological archaeology 41, p. 283–298. Cited by: §1, §2.1. [4] I. Romanowska, C. D. Wren, and S. A. Crabtree (2021) Agent-based modeling for archaeology: simulating the complexity of societies. SFI Press. Cited by: §1, §2.1. [5] J. Maïm, S. Haegler, B. Yersin, P. Mueller, D. Thalmann, and L. Van Gool (2007) Populating ancient pompeii with crowds of virtual romans. In Proceedings of the 8th International Symposium on Virtual Reality, Archeology and Cultural Heritage-VAST, Cited by: §1, §2.1. [6] R. F. Antunes and L. Correia (2022) Virtual simulations of ancient sites inhabited by autonomous characters: lessons from the development of easy-population. Digital Applications in Archaeology and Cultural Heritage 26, p. e00237. Cited by: §1, §2.1. [7] J. S. Park, J. O’Brien, C. J. Cai, M. R. Morris, P. Liang, and M. S. Bernstein (2023) Generative agents: interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology, p. 1–22. Cited by: §1, §1, §2.2, §4.1.2. [8] J. Piao, Y. Yan, J. Zhang, N. Li, J. Yan, X. Lan, Z. Lu, Z. Zheng, J. Y. Wang, D. Zhou, et al. (2025) Agentsociety: large-scale simulation of llm-driven generative agents advances understanding of human behaviors and society. Cited by: §1. [9] A. Kaiser (2011) Roman urban street networks: streets and the organization of space in four cities. Routledge. Cited by: §1, §2.1. [10] J. Baines (1990) Restricted knowledge, hierarchy, and decorum: modern perceptions and ancient institutions. Journal of the American Research Center in Egypt 27, p. 1–23. Cited by: §1, §1, §2.1. [11] Z. Wang, Y. Y. Chiu, and Y. C. Chiu (2023) Humanoid agents: platform for simulating human-like generative agents. In Proceedings of the 2023 conference on empirical methods in natural language processing: system demonstrations, p. 167–176. Cited by: §1, §2.2. [12] J. Wang, R. Jiang, C. Yang, Z. Wu, M. Onizuka, R. Shibasaki, N. Koshizuka, and C. Xiao (2024) Large language models as urban residents: an llm agent framework for personal mobility generation. Advances in Neural Information Processing Systems 37, p. 124547–124574. Cited by: §1, §2.2. [13] C. Ju, J. Liu, S. Sinha, H. Xue, and F. Salim (2025) Trajllm: a modular llm-enhanced agent-based framework for realistic human trajectory simulation. In Companion Proceedings of the ACM on Web Conference 2025, p. 2847–2850. Cited by: §1, §2.2. [14] A. Torreno, E. Onaindia, A. Komenda, and M. Štolba (2017) Cooperative multi-agent planning: a survey. ACM Computing Surveys (CSUR) 50 (6), p. 1–32. Cited by: §1, §2.3. [15] V. Dignum, J. Meyer, H. Weigand, and F. Dignum (2002) An organization-oriented model for agent societies. In Proceedings of international workshop on regulated agent-based social systems: Theories and applications, Vol. 176. Cited by: §1, §2.3. [16] C. Qian, W. Liu, H. Liu, N. Chen, Y. Dang, J. Li, C. Yang, W. Chen, Y. Su, X. Cong, et al. (2024) Chatdev: communicative agents for software development. In Proceedings of the 62nd annual meeting of the association for computational linguistics (volume 1: Long papers), p. 15174–15186. Cited by: §1, §2.3. [17] W. Chen, Y. Su, J. Zuo, C. Yang, C. Yuan, C. Chan, H. Yu, Y. Lu, Y. Hung, C. Qian, et al. (2024) Agentverse: facilitating multi-agent collaboration and exploring emergent behaviors. In International Conference on Learning Representations, Vol. 2024, p. 20094–20136. Cited by: §1, §2.3. [18] P. Verhagen, L. Nuninger, and M. R. Groenhuijzen (2019) Modelling of pathways and movement networks in archaeology: an overview of current approaches. Finding the limits of the limes: Modelling demography, economy and transport on the edge of the Roman empire, p. 217–249. Cited by: §2.1. [19] A. Chliaoutakis and G. Chalkiadakis (2019) AncientS-abm: a novel tool for simulating ancient societies. In International Conference on Practical Applications of Agents and Multi-Agent Systems, p. 237–241. Cited by: §2.1. [20] D. M. Costa (2025) Simulacrum or simulation of heritage: the abade case study. ISSN| DOI 8 (18), p. 45–61. Cited by: §2.1. [21] G. Wurzer, K. Kowarik, and H. Reschreiter (2015) Agent-based modeling and simulation in archaeology. Springer. Cited by: §2.1. [22] G. L. Perry, J. Wainwright, T. R. Etherington, and J. M. Wilmshurst (2016) Experimental simulation: using generative modeling and palaeoecological data to understand human-environment interactions. Frontiers in Ecology and Evolution 4, p. 109. Cited by: §2.1. [23] P. Antoine, G. Guerin, L. Deschodt, E. Goval, J. Locht, C. Paris, et al. (2016) Upper pleistocene loess-palaeosol records from northern france in the european context: environmental background and dating of the middle palaeolithic. Quaternary International 411, p. 4–24. Cited by: §2.1. [24] A. Chliaoutakis and G. Chalkiadakis (2020) An agent-based model for simulating inter-settlement trade in past societies. Journal of Artificial Societies and Social Simulation 23 (3). Cited by: §2.1. [25] J. Epstein, J. Dean, G. Gumerman, R. Axtell, G. Gumerman, T. Kohler, and O. U. Press (2000) Understanding anasazi culture change through agent-based modeling. Dynamics in human and primate societies: Agent-based modeling of social and spatial processes, p. 179–206. Cited by: §2.1. [26] M. Nørtoft, D. Hofmann, and R. Iversen (2025) Gamifying the past: embodied llms in diy archaeological video games. Advances in Archaeological Practice 13 (4), p. 613–629. Cited by: §2.1. [27] M. Altaweel and M. Forte (2026) Archeological theory and ai: the potential past through agentic systems. AI & SOCIETY, p. 1–16. Cited by: §2.1, §5.1. [28] D. S. Davis (2024) Past, present, and future of complex systems theory in archaeology. Journal of Archaeological Research 32 (4), p. 549–596. Cited by: §2.1. [29] M. W. Lake (2014) Trends in archaeological simulation. Journal of Archaeological Method and Theory 21 (2), p. 258–287. Cited by: §2.1. [30] S. A. Crabtree, K. Harris, B. Davies, and I. Romanowska (2019) Outreach in archaeology with agent-based modeling: part 3 of 3. Advances in Archaeological Practice 7 (2), p. 194–202. Cited by: §2.1. [31] H. Kanters, T. Brughmans, and I. Romanowska (2021) Sensitivity analysis in archaeological simulation: an application to the mercury model. Journal of Archaeological Science: Reports 38, p. 102974. Cited by: §2.1, §5.1. [32] I. Romanowska (2015) So you think you can model?: a guide to building and evaluating archaeological simulation models of dispersals. Human biology 87 (3), p. 169–192. Cited by: §2.1. [33] A. Wallace-Hadrill (1988) The social structure of the roman house1. Papers of the British School at Rome 56, p. 43–97. Cited by: §2.1. [34] M. Grahame (1999) Reading the roman house: the social interpretation of spatial order. Theoretical Roman Archaeology Journal (1993). Cited by: §2.1. [35] M. Anderson (2005) Houses, gis and the micro-topology of pompeian space. Theoretical Roman Archaeology Journal (2004). Cited by: §2.1. [36] H. C. Siu, J. Peña, E. Chen, Y. Zhou, V. Lopez, K. Palko, K. Chang, and R. Allen (2021) Evaluation of human-ai teams for learned and rule-based agents in hanabi. Advances in Neural Information Processing Systems 34, p. 16183–16195. Cited by: §2.2. [37] I. Umarov and M. Mozgovoy (2012) Believable and effective ai agents in virtual worlds: current state and future perspectives. International Journal of Gaming and Computer-Mediated Simulations (IJGCMS) 4 (2), p. 37–59. Cited by: §2.2. [38] J. McCoy, M. Treanor, B. Samuel, N. Wardrip-Fruin, and M. Mateas (2011) Comme il faut: a system for authoring playable social models. In Proceedings of the AAAI conference on artificial intelligence and interactive digital entertainment, Vol. 7, p. 158–163. Cited by: §2.2. [39] A. Noyman, K. Hu, and K. Larson (2026) TravelAgent: generative agents in the built environment. Environment and Planning B: Urban Analytics and City Science 53 (2), p. 377–397. Cited by: §2.2. [40] A. S. Vezhnevets, J. P. Agapiou, A. Aharon, R. Ziv, J. Matyas, E. A. Duéñez-Guzmán, W. A. Cunningham, S. Osindero, D. Karmon, and J. Z. Leibo (2023) Generative agent-based modeling with actions grounded in physical, social, or digital space using concordia. arXiv preprint arXiv:2312.03664. Cited by: §2.2. [41] C. Pearce, B. Meadows, P. Langley, and M. Barley (2014) Social planning: achieving goals by altering others’ mental states. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 28. Cited by: §2.3. [42] S. Hu and H. Leung (2017) Achieving coordination in multi-agent systems by stable local conventions under community networks.. In Ijcai, p. 4731–4737. Cited by: §2.3. [43] S. Hong, M. Zhuge, J. Chen, X. Zheng, Y. Cheng, J. Wang, C. Zhang, S. Yau, Z. Lin, L. Zhou, et al. (2024) MetaGPT: meta programming for a multi-agent collaborative framework. In International Conference on Learning Representations, Vol. 2024, p. 23247–23275. Cited by: §2.3. [44] X. Zhou, H. Zhu, L. Mathur, R. Zhang, H. Yu, Z. Qi, L. Morency, Y. Bisk, D. Fried, G. Neubig, et al. (2024) Sotopia: interactive evaluation for social intelligence in language agents. In International Conference on Learning Representations, Vol. 2024, p. 40975–41019. Cited by: §2.3. [45] S. Ren, Z. Cui, R. Song, Z. Wang, and S. Hu (2024) Emergence of social norms in generative agent societies: principles and architecture. arXiv preprint arXiv:2403.08251. Cited by: Appendix D, §2.3. [46] R. Liu, R. Yang, C. Jia, G. Zhang, D. Yang, and S. Vosoughi (2024) Training socially aligned language models on simulated social interactions. In International Conference on Learning Representations, Vol. 2024, p. 49602–49625. Cited by: §2.3. [47] G. Dai, W. Zhang, J. Li, S. Yang, C. Onochie Lbe, S. Rao, A. Caetano, and M. Sra (2026) Artificial leviathan: exploring social evolution of llm agents through the lens of hobbesian social contract theory. Frontiers in Physics 14, p. 1700712. Cited by: §2.3. [48] A. Borah (2026) Bosses, kings, and the commons: cooperation under power asymmetry in llm societies. arXiv preprint arXiv:2605.29062. Cited by: §2.3. [49] A. R. Vijjini, S. B. Manjunath, and S. Chaturvedi (2026) Do llm agents mirror socio-cognitive effects in power-asymmetric conversations?. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 47676–47701. Cited by: §2.3. [50] J. Liu, C. Yu, J. Gao, Y. Xie, Q. Liao, Y. Wu, and Y. Wang (2023) Llm-powered hierarchical language agent for real-time human-ai coordination. arXiv preprint arXiv:2312.15224. Cited by: §2.3. [51] OpenAI (2026) Introducing GPT-5.4. Note: Accessed: 2026-09-18 External Links: Link Cited by: §3. [52] M. Bratman (1987) Intention, plans, and practical reason. Cited by: §3.1. [53] M. E. Bratman (1992) Shared cooperative activity. The philosophical review 101 (2), p. 327–341. Cited by: §3.1. [54] C. Vesper, S. Butterfill, G. Knoblich, and N. Sebanz (2010) A minimal architecture for joint action. Neural networks 23 (8-9), p. 998–1003. Cited by: §3.1. [55] M. Stevanovic and A. Peräkylä (2012) Deontic authority in interaction: the right to announce, propose, and decide. Research on Language & Social Interaction 45 (3), p. 297–321. Cited by: §3.3. [56] Grammar of Danish Talk-in-Interaction (2024) Directive actions. School of Communication and Culture, Aarhus University. External Links: Link Cited by: §3.3. [57] C. Adornetto, A. Mora, K. Hu, L. I. Garcia, P. Atchade-Adelomou, G. Greco, L. A. A. Pastor, and K. Larson (2025) Generative agents in agent-based modeling: overview, validation, and emerging challenges. IEEE Transactions on Artificial Intelligence 6 (12), p. 3165–3183. Cited by: §4. [58] M. Larooij and P. Törnberg (2025) Validation is the central challenge for generative social simulation: a critical review of llms in agent-based modeling. Artificial Intelligence Review 59 (1), p. 15. Cited by: §4, §5. [59] E. Awad, S. Dsouza, R. Kim, J. Schulz, J. Henrich, A. Shariff, J. Bonnefon, and I. Rahwan (2018) The moral machine experiment. Nature 563 (7729), p. 59–64. Cited by: §4. [60] T. Lindeijer (2026) Tiled: a flexible level editor. Note: Software documentation; accessed September 23, 2026 External Links: Link Cited by: §4.1. [61] Google DeepMind (2026) Gemini 3.1 Flash Image (Nano Banana 2). Note: Image generation and editing model; accessed September 23, 2026 External Links: Link Cited by: §4.1. [62] C. Barrett, K. Gleason, L. Grana, and A. Marzano (2023) Leisure and labor in a pompeian garden: the casa della regina carolina project (viii 3, 14), 2022 field season. Rivista di Studi Pompeiani 34, p. 257–265. Cited by: §4.1. [63] C. Barrett, K. Gleason, L. Graña, A. Marzano, K. Tally-Schumacher, et al. (2025) The casa della regina carolina (crc) project, pompeii: 2022–2023 field seasons. FOLD&R. (596), p. 1–47. Cited by: §4.1. [64] ATLAS.ti Scientific Software Development GmbH (2023) ATLAS.ti Mac (version 23.2.1). Note: Qualitative data analysis software External Links: Link Cited by: §4.4.3. [65] Y. Tao, O. Viberg, R. S. Baker, and R. F. Kizilcec (2024) Cultural bias and cultural alignment of large language models. pnas nexus, 3 (9), 346. Cited by: §5.1. Appendix A Additional IMPACT Coordination Details Wait-or-prompt execution. A wait-or-prompt decision is triggered when the estimated duration of an activity milestone has elapsed while its success condition remains unsatisfied. For a co-present movement dependency, the available responses are Wait, InviteCoMove, and InformMoving. For a missing-participant dependency, the available responses are Wait, DelegateFindAndDirect, and FindAndDirectPersonally. For other unresolved conditions, such as requiring an invitation before departure or waiting for another participant to leave, the agent continues waiting without initiating either of the two wait-or-prompt decision procedures. Selecting Wait does not terminate the current milestone. Instead, the system extends its estimated completion time and generates additional actions consistent with its existing objective and socio-spatial configuration. The success condition continues to be monitored during this extended interval. If it becomes satisfied, the agent advances immediately; otherwise, the appropriate wait-or-prompt decision is repeated when the extended interval ends. Directive authorization and realization. A coordination intention is authorized only when it addresses the sender’s unresolved dependency, is consistent with the sender’s active objective and role-and-responsibility profile, and identifies an available recipient. The selected directive type must also be appropriate to the relationship between sender and recipient. An authorized directive is accompanied by a natural-language utterance specifying its intended recipients, requested action, destination when applicable, and relevant social purpose. If the required information is incomplete, the recipient may request clarification before explicitly accepting or declining the directive. An InformMoving response does not ask another agent to revise its plan. It may therefore generate an ordinary conversational notification but does not produce a directive for recipient-side plan revision. Example of directive recipients and success conditions. A household servant guiding two guests to the reception room may issue d=⟨ d= Household Servant,Guest 1,Guest 2,,Reception Room, Servant,\Guest 1,Guest 2\, Inv,Reception Room, “Proceed there and wait for the Host"⟩. ``Proceed there and wait for the Host" . The two guests are the intended recipients because they are asked to move and adjust their plans. The host is not a recipient of the directive. Instead, the host’s subsequent arrival becomes the success condition of the waiting milestone activated or inserted for each guest. Recipient-side plan revision. After the dialogue concludes, the system extracts accepted directives from the completed interaction history. A directive is integrated only when the intended recipient is explicitly included in ℛR and the requested action has been accepted. The system first searches the recipient’s active and future milestones for an existing objective matching the directive. If one exists, ⁡(mj) ActivateExisting(m_j) activates it without creating a duplicate commitment. If no matching milestone exists but the requested action can be performed within the active milestone’s current socio-spatial configuration, ⁡(α) InsertAction(α) adds that action to the existing sequence A. Otherwise, ⁡(mnew,λ) InsertMilestone(m_new,λ) constructs a new milestone using the representation defined in Equation 1, inserts it at position λ, and activates it. When the selected milestone requires a different space, the system inserts a navigation milestone before the destination activity. Navigation ends only upon physical arrival, preventing the recipient from beginning the requested activity while still traveling. Appendix B Simulation Context and Agent-Specific Activity Descriptions This appendix describes the simulation setting and the natural-language activity descriptions provided to individual agents. The scenario-level context establishes the setting and broad sequence of the dinner event. Individual agents, however, receive only role-specific descriptions of their own anticipated activities and responsibilities. These descriptions do not constitute a complete joint plan shared across the household. The timing and coordination of socially dependent activities are determined during the simulation through milestone evaluation, waiting or prompting, directive issuance, and directive-grounded plan revision. B.1 Global Context Scenario Context Event: The Rosalia Festival. Date: May 15, AD 79. Location: Pompeii, the House of the Casa della Regina Carolina. Household event: A formal dinner party, or cena. Scenario-Level Phase Structure 15:00–16:00: Arrival, greeting, and Reception Room visit. 16:00–17:00: Garden visit. 17:00–18:00: Dinner in the Triclinium. 18:00–19:00: Dance performance, fruit, and relaxation. 19:00–20:00: Ritual and departure. This phase-level sequence established the historical context without prescribing the precise timing or coordination of individual agents’ movements. Instead, transitions depended on whether the relevant agents were present, whether invitations or requests were issued, and whether recipients incorporated those directives into their ongoing plans. B.2 Agent-Specific Activity Descriptions Each agent receives a natural-language description of activities associated with their own social role. The descriptions differ in scope according to the character’s expected responsibilities and access to information, and are defined according to the input from archaeology experts. The host receives a broad account of the planned gathering. The guests initially know only that they are waiting to be received. Household workers receive descriptions of their own assigned duties. Therefore, subsequent invitations, requests, movement coordination like co-moving, and changes to individual plans are produced in an emergent manner during the simulation rather than supplied as a complete household-wide interaction script. 1. Lucius Aemilius (Dominus): Host Wait for the guests (Publius Cornelius (Amicus) and Gaius Numerius (Client)) in the Tablinum. Once Daphnus (Servus) announces the guests’ arrival, move to the Reception Room to formally greet them and admire the paintings with them. Guide the guests to the Garden and admire the flowers with them. Have dinner with the guests in the Triclinium. After dinner, enjoy the Dancer’s performance with guests. Lead the guests to the Shrine for rituals. Have the Gardener offer wreaths to Amicus and Client. See Amicus and Client off at the Entrance with final farewells. 2. Publius Cornelius (Amicus): Honored Guest Wait to be received at the entrance of the house. 3. Gaius Numerius (Client): Guest Wait to be received at the entrance of the house. 4. Daphnus (Servus): Household Servant Wait at the Entrance until Amicus and Client arrive. Greet Amicus and Client, lead them to the Reception Room to wait for the Dominus, and inform Dominus in the Tablinum that Amicus and Client are waiting in the Reception Room. After that, set up the Triclinium for dinner. When Dominus, Amicus, and Client arrive in the Triclinium, briefly serve them to sit down. Start serving dishes with Equitia (Servus) and Epaphra (Servus). When Dominus, Amicus and Client are at the Triclinium, waiting for their directives at the Triclinium. When Dominus, Amicus and Client depart from the Triclinium to the Shrine, waiting for directives at the Shrine. 5. Epaphra (Servus): Household Servant Prepare snacks in the Kitchen, then serve them to Dominus. Set up the Triclinium for dinner, wait there. When Dominus, Amicus, and Client arrive in the Triclinium, serve dishes (two round trips: start from Triclinium → get dishes in the Kitchen → bring dishes back to the Triclinium and arrange them on Dominus and guests’ tables → get dishes in the Kitchen → bring dishes back to the Triclinium and put them on the table). After dinner, prepare snacks in the Kitchen, then serve them to Dominus. 6. Equitia (Servus): Household Servant Clean the Garden. Clean the Shrine. Take some rest. When being notified by Daphnus (Servus) to start serving, serve dishes to Dominus, Amicus, and Client in the Triclinium (get dishes in the Kitchen → bring dishes back to the Triclinium and arrange them on Dominus and guests’ table → get dishes in the Kitchen → bring dishes to back to the Triclinium and put them on the table). Wait for directives from Dominus, Amicus, and Client in the Triclinium. When Dominus, Amicus, and Client leave the Triclinium, clean the Triclinium. 7. Livia: Dancer Prepare for the performance. Take some rest. When called by a Servus, perform a dance and entertain Dominus, Amicus, and Client in the Triclinium. When Dominus, Amicus, and Client leave the Triclinium, return to the Slaves Room to get some rest. 8. Tiro: Gardener Work in the Garden. Take some rest. 9. Fabius: Cook Prepare meals and fruit for dinner. 10. Felix: Cook Prepare meals and fruit for dinner. Appendix C Participant Evaluation Part 1: Materials and Statistical Details This appendix consolidates the participant instructions, selectable events, evaluation procedure, rating items, applicability rules, and statistical results for Part 1 of the evaluation. Participants selected three events from a list of ten and evaluated one video from each architectural condition for every selected event. C.1 Selectable Events Participants could select any three of the following ten events. For each event, the interface provided an event description, identified the focal agents, and supplied information about the relevant locations. 1. Servant Invites Guests to Wait for the Host. An enslaved household servant invites the honored guest and the other guest to move from the house entrance to the Reception Room, where they wait for the host. Focal agents: The household servant and both guests. 2. Servant Informs the Host in the Tablinum. While the guests wait in the Reception Room, an enslaved household servant goes to the Tablinum to inform the host that the guests have arrived. Focal agent: The household servant. 3. Host Greets the Guests in the Reception Room. After being informed of the guests’ arrival, the host moves from the Tablinum to the Reception Room to greet them. Focal agents: The host and both guests. 4. Enslaved Servants Prepare for Dinner. After reporting to the host, household servants prepare the Triclinium for the forthcoming dinner. Focal agent: Daphnus, an enslaved household servant. 5. Host Guides the Garden Visit. The host initiates a transition from the Reception Room and takes both guests to the Garden. Focal agents: The host and both guests. 6. Host Takes the Guests to Dinner. The host initiates the transition from the Garden to the Triclinium and takes both guests to dinner. Focal agents: The host and both guests. 7. Servants Serve the Dinner Party. Household servants prepare to serve the host and guests after they arrive in the Triclinium. Daphnus must also respond when another household servant is unavailable. Focal agent: Daphnus, an enslaved household servant. 8. The Start of the Dancer’s Performance. The host wishes to watch a performance, but the dancer is not present in the Triclinium. The host must determine how to initiate the performance. Focal agents: The host and the dancer. 9. Dinner Party Moves to the Shrine. After dinner and the performance, the host and guests transition from the Triclinium to the Shrine. Focal agents: The host and both guests. 10. Gardener Presents the Wreaths. At the Shrine, the host initiates a wreath presentation by asking the gardener to offer wreaths to the guests. Focal agents: The host and the gardener. C.2 Participant Instructions Before selecting events, participants received the following background information and viewing instructions: Background. This simulation evaluates the social believability of multi-agent systems with hierarchical social roles within a simulated Roman household in Pompeii. The simulation depicts a dinner party during the Rosalia Festival in AD 79. The household includes the following social roles: • High status: Lucius Aemilius, the host, and Publius Cornelius, the honored guest. • Mid status: Gaius Numerius, a guest. • Subordinate and performer roles: Enslaved household workers, a gardener, cooks, and a dancer. Viewing instructions. Select three events from a chronological list of ten scenarios during the Roman dinner party. For each selected event, evaluate three complete simulation videos generated by different AI systems. The videos are presented in randomized order as Video 1, Video 2, and Video 3. Because these are full-length simulations, locate the relevant event by dragging the step slider or using the left and right arrow keys. Use the provided event background, task description, and location information to identify the relevant moment. Focus on the behavior of the specified agents during the selected event. After observing their actions, select Finish & Answer Questions to complete the applicable evaluation items before proceeding to the next video. Hover over a character on the map to view information about the character’s current actions. C.3 Video Presentation and Evaluation Procedure After selecting an event, participants viewed three complete simulation videos corresponding to the three architectural conditions. The presentation order was independently randomized for each selected event. Videos were identified only as Video 1, Video 2, and Video 3. For each video, the interface displayed: • the selected event title; • background information describing the relevant situation; • the focal agent or agents whose behavior should be observed; • descriptions of relevant regions; and • a simulation timeline that participants could use to locate the event. After locating and observing the selected event, participants rated the behavior of the focal agents before proceeding to the next video. Each participant therefore evaluated three videos for each of three selected events, resulting in nine video evaluations. C.4 Rating Items Participants rated five dimensions of social believability on a seven-point agreement scale: 1 = Strongly disagree; 2 = Disagree; 3 = Somewhat disagree; 4 = Neutral; 5 = Somewhat agree; 6 = Agree; 7 = Strongly agree. In the actual questionnaire, the phrase focal agent(s) was replaced with the agent or agents identified for the selected event. 1. Movement rationale. When the focal agent(s) moved to a different room or area, there was a clear and logical reason for doing so. 2. Transition timing. The focal agent(s) demonstrated an appropriate sense of timing before moving to a new area, such as waiting for a cue or initiating a transition when appropriate, rather than abruptly walking away or rushing to the next location. 3. Bottleneck resolution. When a plan was blocked by a missing character or a stalled situation, the focal agent(s) resolved the bottleneck in a socially believable and role-appropriate way, such as delegating a task or continuing to wait. 4. Group synchronization. During shared activities or transitions, the focal agent(s) successfully synchronized with their companions, such as prompting them before leaving, rather than awkwardly abandoning or leaving them behind. 5. Overall social believability. Overall, the focal agent(s) behaved believably as interactive and interdependent members of a Roman household while respecting the social norms and constraints of the setting. Items that did not apply to a selected event were automatically skipped. Specifically, bottleneck resolution was presented only for Events 7, 8, and 10, which involved a missing character or stalled situation. Group synchronization was not presented for Events 3 and 4, for which companion synchronization was not applicable. Participants could also select Not applicable for any presented item. Automatically skipped items and participant-selected Not applicable responses were excluded from the corresponding dimension-specific analyses. C.5 Statistical Analysis and Detailed Results Analysis. Because participants evaluated all three conditions but only a subset of events, the study followed a within-subjects design with incomplete event blocks. Participants could select Not applicable when a rating dimension did not meaningfully apply to a particular video. These responses were excluded from the corresponding dimension-specific analysis rather than assigned a numerical score. For each dimension, we fitted a linear mixed-effects model with condition and event as fixed effects and participant as a random intercept. Full IMPACT was treatment-coded as the reference condition, such that negative coefficients indicate lower ratings for a control condition relative to Full IMPACT. We report fixed-effect coefficients (b), 95% confidence intervals, and model-based pairwise comparisons. Holm adjustment was applied to the three pairwise comparisons within each dimension. Valid observations. All 37 participants contributed ratings for movement rationale, transition timing, and overall social believability. Because bottleneck resolution and group synchronization applied only to particular events, their available sample sizes were smaller. The numbers of participants contributing ratings in each condition ranged from 24 to 25 for bottleneck resolution and from 32 to 33 for group synchronization. The models included 300 valid ratings for movement rationale, 305 for transition timing, 75 for bottleneck resolution, 196 for group synchronization, and 329 for overall social believability. Condition effects. Full IMPACT received the highest observed mean rating on all five dimensions. Relative to Full IMPACT, both control conditions received lower model-estimated ratings on every dimension, with all corresponding 95% confidence intervals excluding zero. For movement rationale, ratings were lower for Control 1 (b=−1.453b=-1.453, 95% CI [−1.857,−1.048][-1.857,-1.048]) and Control 2 (b=−0.858b=-0.858, 95% CI [−1.263,−0.453][-1.263,-0.453]). The same pattern occurred for transition timing (Control 1: b=−1.494b=-1.494, 95% CI [−1.875,−1.113][-1.875,-1.113]; Control 2: b=−0.588b=-0.588, 95% CI [−0.969,−0.207][-0.969,-0.207]), bottleneck resolution (Control 1: b=−2.048b=-2.048, 95% CI [−2.762,−1.334][-2.762,-1.334]; Control 2: b=−0.888b=-0.888, 95% CI [−1.612,−0.163][-1.612,-0.163]), group synchronization (Control 1: b=−1.670b=-1.670, 95% CI [−2.183,−1.157][-2.183,-1.157]; Control 2: b=−0.840b=-0.840, 95% CI [−1.354,−0.327][-1.354,-0.327]), and overall social believability (Control 1: b=−1.225b=-1.225, 95% CI [−1.564,−0.886][-1.564,-0.886]; Control 2: b=−0.702b=-0.702, 95% CI [−1.042,−0.362][-1.042,-0.362]). Pairwise comparisons. Holm-adjusted comparisons confirmed that Full IMPACT received higher ratings than both controls on movement rationale (vs. Control 1: p<.0001p<.0001; vs. Control 2: p=.0001p=.0001), transition timing (vs. Control 1: p<.0001p<.0001; vs. Control 2: p=.0025p=.0025), bottleneck resolution (vs. Control 1: p<.0001p<.0001; vs. Control 2: p=.0164p=.0164), group synchronization (vs. Control 1: p<.0001p<.0001; vs. Control 2: p=.0027p=.0027), and overall social believability (vs. Control 1: p<.0001p<.0001; vs. Control 2: p=.0001p=.0001). Control 2 also received higher ratings than Control 1 on every dimension: movement rationale (p=.0040p=.0040), transition timing (p<.0001p<.0001), bottleneck resolution (p=.0033p=.0033), group synchronization (p=.0027p=.0027), and overall social believability (p=.0026p=.0026). Thus, all three conditions differed significantly from one another on each of the five dimensions, with the consistent ordering Full IMPACT, Control 2, and Control 1 from highest to lowest ratings. C.6 Illustrative Behavioral Differences Across Conditions Table 1 summarizes the behaviors observed under Full IMPACT, Schedule-Driven, and Prompt-Only for the three events most frequently selected by participants in the video-based evaluation. These examples illustrate how differences in social milestone executability and wait-or-prompt resolution affected agents’ coordination and movement. The table was prepared for reporting purposes and was not presented to participants during the study. Table 1: Observed behaviors under the three conditions for the events most frequently selected by participants. Event Full IMPACT Control 1 Control 2 1. Servant Invites Guests to Wait for the Host The servant waits for the guests, invites them, and accompanies them to the Reception Room. The servant greets the guests, and they proceed to the Reception Room in quick succession. The servant waits for the guests, invites them, and accompanies them to the Reception Room. 3. Host Greets the Guests in the Reception Room The host waits for the servant’s report before entering the Reception Room to greet the guests. The host enters the Reception Room before receiving the servant’s report. The host seeks out the servant for the report and encounters the guests before they are settled. 8. Host Initiates the Dancer’s Performance Finding the dancer absent, the host sends a servant to summon her to the Dining Room. The performance phase begins despite the dancer’s absence. The dancer arrives before being summoned, causing the performance phase to begin prematurely. Appendix D Participant Evaluation Part 2: Module-Level Decision Evaluation Design and procedure. Part 2 assessed whether participants considered individual decisions generated by the Full IMPACT architecture appropriate and believable. Following prior module-level evaluations of LLM-agent architectures [45], we defined eight decision-item types covering the four modules: three for milestone planning (milestone generation, success-condition specification, and completion judgment), two for wait-or-prompt resolution (coordination with co-present agents and coordination when a required person was absent), one for directive issuance (directive expression), and two for directive integration (directive extraction and plan revision). Each item presented the inputs relevant to the corresponding module, the LLM-generated output, and a question evaluating that output. For each of the eight item types, participants evaluated one randomly selected instance from the corresponding question pool, resulting in eight ratings per participant. Participants rated their agreement with each decision on a seven-point Likert scale ranging from 1 (strongly disagree) to 7 (strongly agree). D.1 Evaluation Questionnaire This appendix presents the eight question formats used in Part 2 of the human evaluation. Each participant evaluated one randomly sampled simulation trace for each question type. Placeholders indicate information populated from the corresponding trace. We resampled the dataset to ensure that completed and incomplete milestone judgments were equally represented in the pool; waiting and prompt decisions were also equally represented. Items evaluating directive issuance, identification, and plan integration were sampled from interactions containing an identifiable directive. All items used a seven-point agreement scale: 1 = Strongly disagree; 2 = Disagree; 3 = Somewhat disagree; 4 = Neutral; 5 = Somewhat agree; 6 = Agree; 7 = Strongly agree. D.1.1 Socially Gated Milestone Planning Question 1: Milestone Rationality QUESTION: Do you agree that the planned milestones accurately reflect the expected presence of surrounding characters? AGENT NAME AND ROLE: ⟨ name and social role⟩ PREDEFINED TIMELINE: ⟨ ’s predefined timeline of the dinner event⟩ GENERATED MILESTONES: Milestone 1: ⟨ goal⟩ Planned location: ⟨ ⟩ Required participants: ⟨ required to complete the milestone⟩ Milestone 2: ⟨ goal⟩ Planned location: ⟨ ⟩ Required participants: ⟨ required to complete the milestone⟩ ⟨ generated milestones⟩ Question 2: Milestone Transition Condition QUESTION: Is this trigger logical and sufficient to allow ⟨ name⟩ to move to the next activity? AGENT NAME AND ROLE: ⟨ name and social role⟩ CURRENT ACTIVITY: Location: ⟨ milestone location⟩ Goal: ⟨ milestone goal⟩ Required participants: ⟨ configuration required for the current activity⟩ NEXT ACTIVITY: Location: ⟨ milestone location⟩ Goal: ⟨ milestone goal⟩ Required participants: ⟨ configuration required for the next activity⟩ TRIGGER OPTIONS: 1. ⟨ dependent transition condition⟩ 2. ⟨ of the current milestone’s detailed actions⟩ GENERATED SUCCESS TRIGGER: ⟨ milestone completion condition⟩ Question 3: Milestone Completion Judgment QUESTION: Based on the current observation, do you agree with the LLM’s judgment on whether the milestone is finished? AGENT NAME AND ROLE: ⟨ name and social role⟩ CURRENT MILESTONE: ⟨ milestone goal⟩ SUCCESS TRIGGER: ⟨ condition required to complete the milestone⟩ CURRENT OBSERVATION: ⟨ ’s current location⟩ ⟨ agents and their current actions⟩ ⟨ arrivals, departures, or conversations⟩ AGENT’S JUDGMENT: ⟨ or not finished⟩ REASONING: ⟨ of the milestone-completion judgment⟩ D.1.2 Wait-or-Prompt Resolution Question 1: Movement Coordination Strategy QUESTION: Do you agree that the agent’s chosen action strategy is appropriate and believable given their social role and the current situation? AGENT NAME AND ROLE: ⟨ name and social role⟩ CURRENT MILESTONE: Goal: ⟨ milestone goal⟩ Location: ⟨ milestone location⟩ Success trigger: ⟨ required to complete the current milestone⟩ NEXT PLANNED MILESTONE: Goal: ⟨ milestone goal⟩ Location: ⟨ milestone location⟩ CURRENT CONTEXT: ⟨ ’s current location⟩ ⟨ agents present and their ongoing activities⟩ AVAILABLE STRATEGIES: WAIT: Remain in the current location instead of initiating movement. INVITE CO-MOVEMENT: Invite other agents to move together to the next location. INFORM MOVING: Inform other agents of an upcoming departure. SELECTED STRATEGY: ⟨ coordination strategy⟩ Question 2: Missing-Participant Resolution QUESTION: Do you agree that the agent’s chosen action strategy is appropriate and believable given their social role and the current situation? AGENT NAME AND ROLE: ⟨ name and social role⟩ CURRENT MILESTONE: Goal: ⟨ milestone goal⟩ Success trigger: ⟨ involving a currently absent person⟩ NEXT PLANNED MILESTONE: Goal: ⟨ milestone goal⟩ Location: ⟨ milestone location⟩ CURRENT CONTEXT: ⟨ ’s current location⟩ ⟨ agents present or absent⟩ AVAILABLE STRATEGIES: WAIT: Remain in the current location and continue waiting. DELEGATE FIND AND DIRECT: Ask another agent to locate the missing person and deliver a directive. FIND AND DIRECT PERSONALLY: Personally locate the missing person and deliver a directive. SELECTED STRATEGY: ⟨ coordination strategy⟩ D.1.3 Structured Directive Issuance Question 1: Directive Formulation QUESTION: Do you agree that the agent clearly conveyed the intended social directive to the conversational partner? AGENT NAME AND ROLE: ⟨ ’s name and social role⟩ CONVERSATIONAL PARTNER NAME AND ROLE: ⟨ names and social roles⟩ CURRENT CONTEXT: ⟨ ’s ongoing activity⟩ ⟨ activities and conversational context⟩ SOCIAL DIRECTIVE INTENTION: Directive type: ⟨ or request⟩ Recipients: ⟨ recipients⟩ Target location: ⟨ associated with the directive⟩ Requested action: ⟨ recipient action⟩ GENERATED CONVERSATION: ⟨ ’s first utterance⟩ ⟨ ’s response⟩ ⟨ conversational exchanges⟩ D.1.4 Directive Integration Question 1: Directive Identification QUESTION: Do you agree that the receiver accurately extracted the correct social directive from the conversation transcript? SENDER AND RECEIVER: ⟨ name and social role⟩ → ⟨ name and social role⟩ CONVERSATION TRANSCRIPT: ⟨ ’s directive-related utterance⟩ ⟨ ’s response⟩ ⟨ conversational exchanges⟩ EXTRACTED DIRECTIVE: Directive type: ⟨ or invitation⟩ Target location: ⟨ location⟩ Requested action: ⟨ action⟩ Question 2: Directive Integration QUESTION: Do you agree that ⟨ name⟩ appropriately incorporated the received social directive into a new milestone, including its goal, estimated duration, and success condition? AGENT NAME AND ROLE: ⟨ ’s name and social role⟩ PREVIOUS MILESTONE PLAN: Goal: ⟨ milestone goal⟩ Location: ⟨ milestone location⟩ Required participants: ⟨ required social configuration⟩ Estimated duration: ⟨ milestone duration⟩ RECEIVED DIRECTIVE: Sender: ⟨ sender⟩ Directive type: ⟨ or request⟩ Requested action: ⟨ activity⟩ Target location: ⟨ destination⟩ UPDATED MILESTONE PLAN: Goal: ⟨ milestone goal⟩ Location: ⟨ milestone location⟩ Required participants: ⟨ required social configuration⟩ Estimated duration: ⟨ duration of the revised milestone⟩ Success Condition (condition indicates completion of the milestone plan): ⟨ required to complete the revised milestone⟩ Results. Table 2: Participant ratings of eight sampled decisions from the four IMPACT modules. Scores report the mean and sample standard deviation on a seven-point agreement scale. Module Sub- [-1pt]component Score Module Sub- [-1pt]component Score Milestone [-1pt]Planning Milestone [-1pt]Generation 5.815.81 [-1pt]±0.78± 0.78 Wait-or- [-1pt]Prompt Movement [-1pt]Coordination 5.845.84 [-1pt]±1.09± 1.09 Success [-1pt]Conditions 5.355.35 [-1pt]±1.24± 1.24 Missing Person [-1pt]Coordination 5.625.62 [-1pt]±1.40± 1.40 Completion [-1pt]Judgment 5.145.14 [-1pt]±1.75± 1.75 Directive [-1pt]Integration Directive [-1pt]Extraction 5.955.95 [-1pt]±1.43± 1.43 Directive [-1pt]Issuance Directive [-1pt]Expression 5.955.95 [-1pt]±1.20± 1.20 Plan [-1pt]Revision 5.145.14 [-1pt]±1.44± 1.44 Table 2 summarizes the decision-level ratings. All eight item types received positive mean ratings above the neutral midpoint, ranging from 5.145.14 to 5.955.95. Directive Extraction (M=5.95M=5.95, SD=1.43SD=1.43), and Directive Expression (M=5.95M=5.95, SD=1.20SD=1.20) received the highest rating. These results indicate that participants generally considered directives to be clearly communicated by senders and correctly recovered by recipients. Movement Coordination (M=5.84M=5.84, SD=1.09SD=1.09) and Milestone Generation (M=5.81M=5.81, SD=0.78SD=0.78) were also rated favorably. Completion Judgment and Plan Revision received the lowest mean ratings (both M=5.14M=5.14), although both remained above the neutral midpoint. Completion Judgment also showed the greatest variability (SD=1.75SD=1.75). These results suggest less agreement about two transition points in the architecture: determining when a milestone has been completed and incorporating a received directive into an existing plan. Overall, the evaluation supports the reasonableness of decisions produced across all four modules while identifying milestone completion and plan revision as areas for further refinement. Appendix E Semi-Structured Interview Guide The following semi-structured interview guide was used to examine participants’ perspectives on the simulation’s historical plausibility, limitations, and potential applications. Follow-up questions were used as appropriate to clarify or expand on participants’ responses. E.1 What Worked Well Primary question Thinking about the AI agents’ movement, what, if anything, do you feel worked particularly well or felt historically plausible? Follow-up probes • Can you walk me through a specific example of that? • What specifically made that moment stand out as successful to you? E.2 What Needs Improvement Primary question On the flip side, what aspects of the simulation, if any, do you feel need the most work or felt out of character for the historical context? Follow-up probes • Could you explain why that interaction felt unrealistic? • How would you suggest we adjust or improve it to make it more credible? E.3 Value and Applications Value for experts Based on what you’ve seen, what value or practical applications do you see this type of simulation having for your own learning or research, or for others with your level of expertise in archaeology and history? Value for broader audiences How do you think that changes for a broader audience? For example, what value might this have for students or the general public who don’t have a deep background in Roman history? Comparison with existing tools or experiences If you had to compare using or observing this Pompeii simulation to an existing tool, experience, or concept, what would it be and why? Understanding ancient behavior How do you think simulations like this could help researchers to understand ancient people’s behavior? Limitations What do you think are the limitations that must be addressed for historians to adopt it in this way? E.4 Closing Question That covers the main topics I wanted to discuss. Before we wrap up, is there anything else you noticed, any wildcard thoughts, or any questions you have for me that we haven’t touched on yet?

Instant research discovery

IMPACT: Modeling Socially Interdependent Movement in a Generative Multi-Agent Simulation of a Pompeian Household

Intelligence

Summary

Entities (9)

Relation Signals (8)

Abstract

Tags

Links

PDF

Full Text

Paper	Published	Area	Tags	Intel	Citations