Methodological Approaches to Studying Group Behavior: From Human Interaction to Agent-Based Models

Human–Human Interaction with Controlled Agents

Understanding group behavior often begins with controlled studies in which real human participants interact under experimentally designed conditions. One promising yet underutilized approach in this domain is Human–Human Interaction with Human-Controlled Agents. Drawing inspiration from the architecture of online multiplayer gaming environments, this approach enables the construction of real-time, interactive platforms where multiple participants control avatars or agents within a shared digital or virtual space (Kyrlitsias & Michael-Grigoriou, 2022; Molyneux et al., 2015; Wu et al., 2025). These agents engage in a variety of coordinated or competitive behaviors—such as following, evading, navigating obstacles, or engaging in simulated conflict—thereby simulating complex social environments that mirror real-world group dynamics.

In this setup, participants may be tasked with interacting alongside or against a group of “avatars”—human-controlled agents who follow predetermined behavioral scripts or protocols. For example, in a simulated fire-escape scenario, each participant sat in front of a computer and viewed the virtual environment from a first-person perspective, which included other participants represented as avatars. Using input devices such as a keyboard and mouse, participants could freely navigate the environment. This setup allows researchers to observe how individuals adapt to or conform with the majority's escape choices. Moussaïd et al. (2016) employed such a platform to investigate crowd behavior in high-risk evacuation situations involving real human participants. In fact, experiments based on economic games can also be considered part of this approach. Although no explicit avatars are present in these games, participants make decisions within predefined rule structures and payoff settings, effectively functioning as controlled agents (Fehr & Gächter, 2002; Rand et al., 2011). This approach of human-human interaction with human-controlled agents allows researchers to explore how an individual's cognition, perception, and decision-making are influenced by, adapt to, and in turn shape the surrounding group behavior. It opens the door to questions such as: How do individuals conform to or diverge from observed group norms? What types of social cues are most influential in shifting behavior? How do moment-to-moment micro-interactions give rise to emergent group-level phenomena?

Unlike traditional confederate-based paradigms—typically conducted in face-to-face laboratory settings, such as Asch's conformity experiments—human-controlled agents are usually implemented in digital platforms or virtual environments. This digital format offers key advantages: interactions can be precisely timed, spatially constrained, and manipulated in context-specific ways, all while maintaining the illusion of natural social presence. Each “avatar” can be remotely directed to display specific behaviors—such as searching for rewards—allowing researchers to examine how subtle cues affect the target participant. For example, Deffner et al. (2024) conducted a naturalistic immersive-reality experiment in which groups of participants searched for rewards across different environments. By manipulating resource distribution (concentrated vs. distributed) and incentive structure (group vs. individual rewards), they found that group-level incentives reduced individuals’ responsiveness to social information and encouraged greater behavioral selectivity over time.

Critically, this approach enables causal inferences about how local interactions contribute to emergent group behavior, as specific behaviors can be systematically selected, controlled, and reproduced across trials. A compelling example of this approach is provided by Nalepka and colleaegues (2017), who investigated multiagent coordination using a digital shepherding task. In this study, pairs of participants were tasked with corralling virtual sheep toward the center of a shared game field. Initially, most participant pairs adopted a complementary “search-and-recover” strategy, in which each person primarily managed sheep on their respective side of the field. However, over time, many pairs spontaneously transitioned to a more effective coordination pattern: a coupled oscillatory containment strategy, wherein both participants moved synchronously around the herd in a dynamic, wave-like motion. These findings demonstrated how simple local interaction rules—such as following the nearest neighbor or avoiding collisions—could give rise to complex emergent patterns at the group level.

This approach has limited capacity to capture truly emergent dynamics, as the “group” consists of controlled agents. To address this, findings are often paired with sensor-based tracking of real group interactions to capture nuanced responses. Nonetheless, human–human interaction with human-controlled agents remains a flexible and powerful approach for studying group behavior under semi-naturalistic yet controlled conditions.

Human–Agent Interaction with Algorithm-Driven Action

To enable more precise manipulation of others’ actions, an increasingly prominent approach involves real humans interacting with algorithm-driven agents. Unlike human-controlled avatars, these agents operate according to preprogrammed rules or machine learning models that simulate social behaviors. This allows for fine-tuned experimental control, particularly over micro-level actions that are often critical to social coordination (Cross & Ramsey, 2021; Jiang et al., 2025). By abstracting away human variability, algorithm-driven agents can be configured with key parameters to isolate and test specific components of social behavior. For example, in studies of social conformity using a virtual fire-escape scenario, a participant controls an avatar via keyboard to select an exit route, while the behaviors of two peer avatars are algorithmically controlled to either choose the same exit or a different one (Yin et al., 2024). This setup enables researchers to precisely manipulate the majority's behavior, define how the majority is formed, and systematically examine conformity effects.

A major strength of this approach is its capacity to deliver rigorous, repeatable interventions while preserving the interactive richness of human social interaction. These agents can be programmed to exhibit emotional expressions, follow social norms, or simulate common cognitive biases. For instance, Hertz et al. (2025) introduced a multiagent reinforcement learning (MARL) approach, which builds on modern artificial intelligence techniques, and provides new avenues to model complex social worlds, while preserving more of their characteristics, and allowing them to capture a variety of social phenomena. Participants may not always recognize the artificial nature of these agents—or, even when aware, still respond to them as if they were genuine social partners. This makes it possible to engage real psychological mechanisms in a highly controlled setting, bridging the gap between experimental precision and ecological relevance. Furthermore, as contemporary social groups increasingly include artificial agents alongside humans, this framework offers critical insights into the evolving dynamics of human-agent collectives and the emerging principles of group behavior in mixed-agent environments.

This approach has been used to investigate how interactive behavior is shaped by subtle social cues (Gao et al., 2010), how societies acquire and sustain social norms (Köster et al., 2022), and how intergroup bias emerges and persists (Nafcha & Hertz, 2024), among others. For example, by populating a foraging environment with artificial agents that must learn to avoid poisonous berries—and introducing a taboo that punishes the consumption of harmless ones—researchers showed that normative behavior develops through a sequence of learned skills, with rule compliance building on prior enforcement by others (Köster et al., 2022). Using algorithmically controlled confederates, Yin et al. (2024) demonstrated that individuals are more likely to follow the majority's choice when selecting from available options than when forced to make a unanimous decision.

The advancement of virtual reality (VR) and robotics has greatly expanded the potential of this paradigm by enabling ecologically valid, scalable, and cost-effective experimentation (Vasser & Aru, 2020). VR, in particular, allows researchers to construct highly immersive and controllable environments that approximate real-world social contexts while preserving experimental precision (Pan & Hamilton, 2018). The use of humanlike avatars further enhances ecological validity by eliciting natural social responses from participants. Through dynamic, interactive scenes, VR makes it possible to examine complex group phenomena such as crowd movement, evacuation dynamics, coordination under uncertainty, and the diffusion of social norms. For example, Zhou et al. (2019) developed a social interaction field model using VR, revealing a front–back asymmetry in social interactions. Moreover, VR platforms can systematically manipulate spatial layouts, perceptual cues, or incentive structures, providing insights into how environmental design influences emergent group patterns. With the addition of wearable motion tracking and physiological monitoring, VR also allows for the simultaneous measurement of fine-grained behavioral and cognitive responses, enriching the analysis of micro-to-macro links in group behavior.

Robotics provides a complementary pathway by offering physical embodiments of agents that mimic humanlike actions in shared environments (Belpaeme et al., 2018). Robots can be programmed to perform consistent, repeatable behaviors, serving as scientific instruments rather than just social partners (Bärmann et al., 2024). This enables researchers to isolate the impact of specific behavioral features—such as gaze direction, movement synchrony, or response delays—on group processes like leadership emergence, role differentiation, or collective decision-making. Unlike VR avatars, robots provide tangible, embodied presence, which can shape participants’ psychological responses in unique ways, such as heightening perceptions of authority, trust, or even competition (Pan & Steed, 2016). For example, in a tablet-based collaborative game with three participants, groups paired with a robot that made vulnerable statements engaged in substantially more conversation, distributed speaking turns more evenly, and evaluated their group more positively compared to groups with a robot that made only neutral or no statements at the end of each round (Traeger et al., 2020).

However, this approach is not without limitations. The artificial nature of the agent's behavior may limit external validity, especially if participants detect the manipulation. Thus, hybrid designs that combine real-time behavioral adaptation (e.g., reinforcement learning) with humanlike algorithms offer a promising middle ground.

Agent–Agent Interaction with Simulation

Group behavior evolves dynamically over time, and certain emergent properties—such as social conventions, cultural norms—often require large-scale temporal frameworks to be observed. Human-participant approaches are inherently limited in capturing such long-term dynamics due to practical constraints and ethical boundaries. At the far end of the experimental spectrum lies agent-agent interaction, typically implemented through computational simulations. In these models, autonomous agents are equipped with human-like attributes—such as cognitive biases, memory, and social preferences. Agent-based modeling (ABM) enables researchers to systematically manipulate agent-level rules to investigate how complex group phenomena emerge over extended timescales, including scenarios that would be ethically impermissible with real participants (Goldstone & Janssen, 2005; Luo et al., 2008; Smith & Conrey, 2007). For instance, in a simulated fire evacuation task, each agent is programmed to escape during an emergency and select routes based on predefined parameters—such as humanlike knowledge of predictable spatial changes relevant to fire safety—allowing researchers to observe how conformity emerges and dynamically changes over time (Tan et al., 2015).

Since the introduction of segregation models, agent-based modeling (ABM) has enabled researchers to iteratively test hypotheses across diverse simulation parameters, revealing robust patterns and delineating boundary conditions for various group dynamics. Simulations also offer unparalleled scalability—thousands of agents can be run concurrently—providing a high–throughput platform for generating hypotheses and rigorously testing theories. ABM has been applied to investigate phenomena such as group formation, norm emergence, and crowd behavior. Gray et al. (2014) employed an ABM grounded in just two principles—reciprocity and transitivity—and demonstrated that agents, despite lacking any shared identity, nonetheless developed strong in–group cooperation and out–group defection. Similarly, Ye et al. (2021) incorporated inertia and trend–seeking mechanisms into their model to reproduce the macroscopic delay and explosive adoption patterns commonly observed in real–-world social diffusion processes.

Recent advances in large language models (LLMs) have substantially expanded the scope of agent–agent interaction studies by enabling agents with flexible, context-sensitive reasoning capacities. Unlike traditional rule-based or reinforcement learning agents that depend on fixed strategies, LLM-driven agents can dynamically interpret instructions, attend to contextual cues, and generate adaptive responses without relying on hard-coded behavioral scripts (Gao et al., 2024). This flexibility allows them to engage in complex social behaviors such as negotiation, role-taking, norm learning, and even cultural transmission, making their group-level dynamics more comparable to those observed in human collectives (Abdurahman et al., 2024). They can also produce language-based justifications, explanations, or persuasive arguments—introducing communicative mechanisms often absent from classical agent-based models (Gao et al., 2024). This opens avenues to study not only what group-level outcomes emerge (e.g., consensus, polarization, cooperation) but also how these outcomes are linguistically constructed and sustained. Moreover, platforms integrating multiple LLM agents—each with distinct roles, goals, or biases—have demonstrated spontaneous emergence of social phenomena such as group norms, hierarchies, and even proto-cultures (e.g., Acerbi & Stubbersfield, 2023; Ashery et al., 2025; Park et al., 2023). For example, Ashery et al. (2025) implemented an LLM-based naming game in which agents drew on accumulated memory of past choices to anticipate their partners’ responses. This pairwise coordination process led to collective biases—such as stable naming conventions—that could not be inferred from isolated agent behavior. By systematically manipulating agent parameters (e.g., memory span, social preference, susceptibility to influence), researchers can now test theories of group psychology at scales and timescales unfeasible with human participants alone.

Ecological exploration of group behavior through agent–agent simulations depends heavily on the algorithms that govern each agent. While LLM-based agents offer a powerful means to simulate collective actions, they often lack the rich personal characteristics and lived experiences of real humans. To enhance realism and predictive validity, these simulations should be augmented with empirical human data—such as behavioral traces, cognitive profiles, and social network information—thereby grounding algorithmic agents in authentic human diversity and complexity.