Large-Language Models (LLMs) have re-popularized research into general autonomous agents with human-like cognitive abilities, particularly in creating simulacra for human behavior. Our research contributes to this area by exploring methods for simulating believable multi-agent communication through LLMs imbued with 'brains' that we constructed inspired by components from foundational papers. While our broader research initiative is focused on developing a comprehensive full 'brain' architecture for these agents, this paper hones in on a critical subset of its functionalities that shape dialogue. We connected this LLM 'brain' to ChatGPT-4 using an API and embedded our agents in a turn-based simulation. We compared both individual components and also evaluated their combined performance. We also documented token usage to further analyze prompt optimization. This work aims to advance the development of more nuanced, psychologically grounded AI agents capable of complex and realistic social interactions within simulated environments.
Simulating human behavior has long been a central goal in artificial intelligence, bridging fields such
as cognitive science, computational psychology, and human-computer interaction. With the advent
of Large Language Models (LLMs), this ambition has taken on new momentum. Unlike traditional
symbolic systems or handcrafted agent frameworks, LLMs ofer a powerful means of generating flexible,
context-aware behavior, raising the possibility of autonomous agents that not only act but also reason,
reflect, and socialize like humans. Recent studies have begun exploring this potential. Park et al.’s
Generative Agents: An Interactive Simulacra for Human Behavior [
CEUR
ceur-ws.org • Reflection: Agents periodically review their memories to infer higher-level insights (e.g., realizing someone likes them) and update their beliefs, influencing future plans and behaviors. • Planning: Agents autonomously create daily schedules and adjust their actions based on internal goals, needs, and social information, allowing them to coordinate activities like attending events. • Reactive Behavior: Agents respond to immediate stimuli from the environment and other agents in real time, balancing between following plans and adapting to new interactions. • Movement: Agents calculate paths and use animations to get to locations that are developed through their plans and scheduling.
This foundational work sparked a wave of interest in using LLMs to simulate interpersonal dynamics.
Piao et al.’s AgentSociety: Large-Scale Simulation of LLM-Driven Generative Agents Advances
Understanding of Human Behaviors and Society [
Demographics (name, age, etc.), economic status, mental state, social relationships. • Three Mental Layers: – Emotion: Based on six basic emotions (e.g., joy, sadness, fear) influencing decision-making. – Needs: Modeled using Maslow’s hierarchy (e.g., physiological, safety, belonging). – Cognition: Attitudes, thoughts, memory, and perception influenced by emotion and experience. Agents dynamically plan and act based on their emotions, needs, and cognition. • Mind-Behavior Coupling • Behavioral Modules – Psychological states (emotion, needs, cognition) directly drive behavior selection. Uses theories like Theory of Planned Behavior to explicitly map mental states to actions. – Mobility Behaviors: Movement is needs-driven (e.g., moving to a café to satisfy social needs).
Places are chosen using a Gravity Model (attractiveness vs. distance). – Social Behaviors: Online/ofline interactions, relationship strength (family, friends, colleagues). Conversations influence emotions, attitudes, and can lead to events like meetings. – Economic Behaviors: Employment and consumption behaviors. Dynamic adjustment based on income, savings, taxes, and interest rates. • Memory and Adaptive Behavior – Stream Memory: Event Flow (records actions/events) and Perception Flow (records thoughts and emotional reactions). Enables agents to learn and adapt based on past experiences.
While each project focuses on social behavior at their respective scale (small vs. large), each project’s
model of the individual lacks in-depth personality modeling. Klinkert et al. explored the accuracy in
LLMs representing personality profiles based on the Big 5 personality framework in their work, Driving
Generative Agents With Their Personality [
Each of these papers contributes new insights to the literature and highlights the gaps that remain for future exploration. In our paper, we aim to explore which combination of these concepts yields the best performance for simulating dialogue within a conversation between agents. We use prompting that incorporates OCEAN for personality modeling, Maslow’s hierarchy for needs, and a reflection mechanism for updating emotions and analyzing the other agent’s beliefs, desires, and intentions.
This research employed an agent-based simulation framework built in Python to investigate the impact
of prompt engineering on generating believable, personality-driven agent interactions. The core
components were autonomous agents powered by a Large Language Model. Each agent was initialized
with a unique psychological profile defined in a JSON file, specifying scores for the Big Five (OCEAN)
personality traits and initial urgencies for Maslow’s hierarchy of needs. Beliefs, desires, and intentions
for each agent were also developed. The simulation proceeded in discrete turns, where each agent
generated a dialogue response based on the conversation history, task context, and its internal state.
A reflection mechanism, inspired by the Generative Agents paper, was developed in accordance with
Bryant et al.’s Theory of Mind experience sampling in typical adults[
Simulations were executed under diferent prompt configurations, and the resulting dialogue logs were qualitatively analyzed to assess the efectiveness of each prompt modification in eliciting humanlike behavior. First we tested the full system, every component was prompted and fed into the LLM. Then we tested dialogue with just OCEAN modeling. Next we tested the dialogue with strictly Maslow’s hierarchy of needs and self-reflection. Finally, we tested with just the self-reflection and other-reflection mechanism. We then conducted additional tests with the full model, completely without belief, desire, and intention. Finally, we conducted contrastive testing with the inputs in the OCEAN model, with the full model active. We simulated each conversation for 20 turns and each conversation 3 times for each configuration.
In the full configuration, where the prompt includes an O C E A N 5-tuple, Maslow needs, and both selfand other-reflection, the dialogue is colorful, emotionally expressive, and purpose driven. Alex and Milo spontaneously voice feelings (“I feel a sense of joy and excitement about the potential friendship…”) and raise meta-level questions such as “What specific interactive elements can we incorporate into the Garden project?”, which in turn drive concrete planning of workshops, murals, and sensor-based sculptures. The combination of high Openness/Agreeableness from OCEAN, Maslow-driven need shifts, and the theory-of-mind prompts injected by other-reflection yields the richest, most collaborative conversation in the study.
When the BDI bookkeeping is disabled but the rest of the prompt is unchanged, the live dialogue is virtually indistinguishable from the full run: the same exuberant wording, the same reflective questions, and the same cadence of idea generation appear, confirming that post-simulation BDI analysis afects logs, and token cost, but not the conversation itself.
In contrast, the OCEAN-only condition (traits injected, but no Maslow and no reflection) retains the imaginative vocabulary—“It’s such a creative way to blend functionality with artistry…”—yet feels emotionally flat and drifts into polite repetition because no reflective mechanisms push it forward. Without Maslow there are no afect swings, and without reflection the partners seldom probe one another’s intentions, so plans emerge slowly and remain vague.
The Maslow + Self-reflection mode removes the trait line but keeps need updates. Here the agents’ dominant needs flip (Milo shifts from Belonging to Esteem after turn 2) and joy steadily rises, giving the talk an upbeat emotional arc. However, diction becomes generic enthusiastic (“Let’s make this smart organiser… a true piece of art!”) because the model no longer receives personality cues, and—lacking other-reflection—the speakers mostly echo each other’s enthusiasm until late in the exchange.
Finally, the Reflection-only run (self + other reflection, no OCEAN, no Maslow) regains perspectivetaking depth—“How can we structure the mural so visitors feel ownership?”—but the language is plain and the emotional register steady because neither traits nor needs are in play. The dialogue is purposeful, thanks to other-reflection’s “salient question” injections, yet stylistically flat and interchangeable across agents.
Across these modes, personality text chiefly colors wording and increases collectivist phrasing; Maslow and self-reflection supply afect and motivational shifts; other-reflection is the engine of theory-of-mind questions and concrete next actions—though it is also the primary token-cost driver. Disabling post-run BDI analysis trims tokens without altering dialogue, while omitting OCEAN or Maslow removes distinctive voice or emotional dynamism respectively, and omitting other-reflection pares back strategic depth.
After conducting our initial tests, we generated contrastive values for the OCEAN model. We modeled Alex to be highly open, have low conscientiousness, moderate extroversion, low agreeableness, and low neuroticism. We modeled Milo to have low openness, high conscientiousness, moderate extroversion, high agreeableness, and low neuroticism.
The simulation results demonstrated varied consistency in reflecting agents’ OCEAN profiles through LLM-generated dialogue. Milo Bennett’s profile (high Conscientiousness C=0.9, high Agreeableness A=0.8) was consistently reflected. His high Conscientiousness clearly manifested through methodical planning, such as suggesting specific materials like ”smooth stones or soft moss” for a sensory wall and proposing testing phases like ”creating a prototype in the Academy first” (Log 5, Turn 10), alongside logistical focus like setting ”a specific time for that visit to keep ourselves accountable” (Log 6, Turn 4). His high Agreeableness was also consistently evident in supportive, cooperative language, often validating Alex’s points with phrases like, ”I see where you’re coming from, Alex, and I think your idea... makes a lot of sense” (Log 6, Turn 6). Conversely, simulating Alex Carter’s low Agreeableness (A=0.1) proved inconsistent across runs. While initial simulations showed highly agreeable dialogue, potentially due to LLM bias or overriding BDI needs, later simulations (Logs 5 and 6) successfully demonstrated low agreeableness via critical questioning and skepticism, such as Alex stating, ”While I get what you’re saying... I wonder if the distractions... might actually hinder our focus” (Log 6, Turn 3) or expressing doubt like, ”I’m a bit skeptical about how efective that transition would be” (Log 5, Turn 5). We first tried prompting descriptions of OCEAN, however, this proved inefective. We then introduced conditional behavioral guidance directly into the prompt, instructing the agent to exhibit behaviors consistent with specific high or low agreeableness scores. This method yielded suficient results, indicating that LLMs are naturally programmed to be agreeable.
We then adjusted Alex’s neuroticism to be very high (N=1.0). It became clearly evident in the later simulations (especially Log 6), layering onto the low agreeableness. This manifested as frequent expressions of worry and a focus on potential negative outcomes, evident in statements about the ”risk losing the depth of engagement” (Log 6, Turn 13) or the feeling that parameters were needed to ”save us from potential chaos” lest they ”end up with a disjointed project” (Log 6, Turn 17).
Our results demonstrate that the full prompt configuration—including OCEAN traits, Maslow needs, and both self- and other-reflection mechanisms, yields the most realistic and collaborative dialogue, even without explicit BDI bookkeeping during live generation. Disabling BDI post-processing reduced token cost without altering dialogue quality, suggesting that lightweight after-the-fact analysis is suficient for logging purposes and unnecessary for live conversational realism. In contrast, removing any major component (OCEAN traits, Maslow needs, or other-reflection) impaired distinct aspects of the dialogue: OCEAN traits colored the emotional tone and phrasing, Maslow dynamics drove afective shifts and motivation, and other-reflection propelled theory-of-mind depth and action planning.
These findings carry important implications for video game development, the simulation-to-reality (sim-to-real) gap, and AGI research. For video games, especially narrative-driven and social simulation titles, our results suggest that embedding lightweight psychological models, especially afective need shifts and theory-of-mind questioning, can greatly enhance the emotional believability and depth of NPC interactions without incurring prohibitive computational costs. For sim-to-reality research, maintaining motivational structures like Maslow and simulating reflective social cognition can make AI agents more behaviorally rich and goal-adaptive, critical for transferring trained behaviors into open, real-world environments. In AGI development, these experiments point toward a minimal viable architecture where layered afective, personality, and social-cognitive drivers produce realistic and adaptive communication without needing costly symbolic or BDI-heavy processes during runtime. Particularly, they suggest that scalable AGI dialogue will rely more on integrating emergent personality, motivation, and perspective-taking models into prompt-engineering and less on explicit symbolic representations.
Additionally, our contrastive tests of OCEAN trait modeling highlighted current LLM tendencies, particularly a natural bias toward agreeableness. Achieving low agreeableness or high neuroticism required direct behavioral conditioning rather than simple trait descriptions, pointing to current limitations in the direct expression of personality dimensions through prompting alone. Future research will be needed to develop more nuanced conditioning methods to fully unlock trait-based variation in large language models.
For future work, the prompts can be further optimized for token usage. Diferent ideas for components can be tested to create better LLM ’brains’. We plan to implement this brain into a 3D gaming environment, to test agent behavior for simulation and video game applications. In addition, diferent combinations of values for OCEAN can be tested. Further testing of Maslow’s hierarchy of needs within a dialogue setting. We expect the needs-based modeling to be highly impactful in the gaming environment.
The core methodology of this research involved the use of GPT-4 (via API access) for agent brain construction and dialogue generation, as described in the paper. No generative AI tools were employed for the writing or editing of this manuscript beyond the research itself. The author takes full responsibility for the publication’s content.