IMPACT: Modeling Socially Interdependent Movement in a Generative Multi-Agent Simulation of a Pompeian Household
Authors: Tianqi Liu, Nayoung Kim, Julia Sebastien, Kathryn Gleason, Caitlín Eilís Barrett, Andrea Stevenson Won
Organizations: Department of Information Science, Cornell University, Ithaca, NY 14853, USA · Department of Communication, Cornell University, Ithaca, NY 14853, USA · Department of Landscape Architecture, Cornell University, Ithaca, NY 14853, USA · Department of Classics, Cornell University, Ithaca, NY 14853, USA
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.
Figures & tables
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.
Figure 2: Reconstruction of the Casa della Regina Carolina simulation environment. Spatial region functions informed by archaeologists.
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.
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 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.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
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
Table 1: Observed behaviors under the three conditions for the events most frequently selected by participants.
Module
Sub- [-1pt] component
Score
Module
Sub- [-1pt] component
Score
Milestone [-1pt]Planning
Milestone [-1pt]Generation
5.81 [-1pt] ±0.78
Wait-or- [-1pt]Prompt
Movement [-1pt]Coordination
5.84 [-1pt] ±1.09
Success [-1pt]Conditions
5.35 [-1pt] ±1.24
Missing Person [-1pt]Coordination
5.62 [-1pt] ±1.40
Completion [-1pt]Judgment
5.14 [-1pt] ±1.75
Directive [-1pt]Integration
Directive [-1pt]Extraction
5.95 [-1pt] ±1.43
Directive [-1pt]Issuance
Directive [-1pt]Expression
5.95 [-1pt] ±1.20
Plan [-1pt]Revision
5.14 [-1pt] ±1.44
Appendix
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.
Studies attempting to simulate human behavior with Silicon Societies grow in numbers while LLM-only social networks have started appearing outside of controlled settings. However, the design space of these networks remains under-studied, which contributes to a gap in validating model realism. To enable future works to make more informed design decisions, we perform a systematic analysis of the consequences and interactions of key design choices in simulated social networks, including the choice of base model used to model individual agents, and how they are connected to each other. Using surveys as a proxy for agent opinions, our findings suggest that the geometry of the design space is non-trivial, with some parameters behaving in additive ways while others display more complex interactions. In particular, the choice of the base LLM is the most important variable impacting the simulation outcomes.
Aurélien Bück-Kaeffer, Sneheel Sarangi, Maximilian Puelma Touzel +3
McGill University · Mila - Quebec Artificial Intelligence Institute · Ubisoft La Forge +2
Large language models can produce locally plausible social turns, but fluent next-turn generation is not enough for social simulation. Human encounters such as restaurant lunches and hotel check-ins are bounded social episodes with roles, scripts, material state, obligations, commitments, timing, and closure conditions. We present EpisodeSim, a hybrid LLM-agent architecture that represents classic-AI structures as natural-language control state interpreted by LLM calls. A World Master maintains shared reality, constructs scenes, adjudicates proposed actions, tracks effects and obligations, and controls closure. Experiments with small qualitative ablations on two held-out settings support a design claim: LLM fluency supplies local texture, but coherent social simulation benefits from persistent classic-AI-style scaffolding that organizes behavior over time.
LLM-agent simulations make natural-language social scenarios easy to instantiate, but their outputs can be overread as predictions and are often difficult to compare with explicit social dynamics. We present AgoraSim, a hybrid agent-based modeling framework for scenario-oriented social reaction analysis. AgoraSim resolves textual or multimodal artifacts into editable ABM configurations, runs ratio-controlled populations that mix LLM, vision-language, custom-endpoint, random, and classical agents, and compares the same scenario against matched classical reference dynamics. All agents emit a shared structured decision object, enabling common action spaces, interaction protocols, metrics, and audit records. Exposed through a local UI, Python SDK/CLI, and REST API, AgoraSim helps users inspect scenario trajectories, compare modeling assumptions, and identify cases that warrant empirical validation.
Chung-Chi Chen
Human-Agent Ally Lab (HAA Lab) National Institute of Informatics, Japan