TLDR: The Psychological-mechanism Agent (PSYA) is a novel AI framework designed to simulate human behavior more authentically by integrating complex emotional, cognitive, and action mechanisms. Based on the Cognitive Triangle, PSYA features modules for multi-layered emotions (short, medium, long-term), goal-directed and spontaneous thinking, and optimized actions driven by needs and feelings. The framework successfully replicates human behaviors in daily life scenarios and classic psychological experiments like learned helplessness and diffusion of responsibility, offering a promising tool for psychological research and realistic agent simulations.
Generative agents have made significant strides in simulating human behavior, but often simplify emotional modeling and focus on specific tasks, which limits the authenticity of their simulations. To address these limitations, researchers Qing Dong, Pengyuan Liu, Dong Yu, and Chen Kang from Beijing Language and Culture University have proposed a novel framework called the Psychological-mechanism Agent (PSYA).
The PSYA framework is built upon the Cognitive Triangle, a psychological concept that emphasizes the interplay of Feeling, Thought, and Action. This design aims to create a more accurate and nuanced simulation of human behavior, moving beyond simple task execution to encompass the complexities of human emotion and cognition.
The Core Modules of PSYA
The PSYA framework consists of three interconnected core modules:
The Feeling module is designed to simulate emotions across different timeframes: short-term, medium-term, and long-term. It uses a layered model of affect, allowing agents to experience immediate emotional responses, prolonged moods, and consistent personality traits. This means an agent might feel a quick burst of surprise, carry a mood of anxiety for days, or consistently exhibit traits like extraversion, making their emotional responses more natural and believable.
The Thought module is based on the Triple Network Model, a concept from neuroscience. It supports both goal-directed thinking and spontaneous, aimless thinking. The Central Executive Network (CEN) handles purposeful tasks like planning, reflection, and decision-making. The Default Mode Network (DMN) simulates mind-wandering, scenario simulation, and self-social cognition, allowing agents to daydream, recall memories, or ponder their self-image. The Salience Network (SN) acts as a switch, determining which thinking mode the agent should engage in based on the context, even introducing random disturbances to make thinking patterns more human-like.
The Action module optimizes agent behavior by integrating emotions, needs, and plans. It determines the agent’s next action, prioritizing tasks based on basic needs (like hunger), emotional states, and task importance. This module also includes a sophisticated conversation trigger mechanism and a “stranger system” that allows agents to collect information about unfamiliar individuals before deciding to interact, enhancing the realism of social interactions.
Also Read:
- Navigating the AI Frontier: Large Language Models in Social Simulation
- Navigating Urban Crowds: A New AI Approach for Safer Autonomous Vehicles
Evaluating PSYA’s Effectiveness
To evaluate the PSYA framework, the researchers conducted both daily life simulations and replicated classic psychological experiments. They extended evaluation metrics to consider self-influence (decisions influenced solely by the agent), one-influence (behavior influenced by a single other agent), and group-influence (behavior influenced by multiple others).
In daily life simulations, PSYA agents demonstrated more natural emotions and diverse, consistent, and credible behaviors compared to standard generative agents. For instance, agents with the emotional module showed prolonged negative emotions after a failure, mirroring human responses, while those with the mind-wandering module exhibited more varied and creative activities to soothe emotions.
The framework successfully replicated outcomes from five classic psychological experiments:
- Learned Helplessness: PSYA agents, particularly with the emotional and self-social cognition modules, showed reduced motivation and problem-solving ability after experiencing uncontrollable negative events, similar to humans.
- Cognitive Dissonance: While initial simulations struggled, modifications to prompts and the introduction of a value system allowed PSYA to replicate how agents adjust their opinions to align with their actions, especially when external justification is low.
- Foot-in-the-Door Effect: PSYA agents were more likely to agree to a larger request after first complying with a smaller one, demonstrating the influence of prior commitment and self-image.
- Diffusion of Responsibility: In simulated emergencies, PSYA agents showed a decrease in the likelihood of taking responsibility as the group size increased, accurately reflecting human behavior in bystander effect scenarios.
- Social Exclusion: Agents experiencing ostracism reported lower levels of belonging, self-esteem, control, and meaningful existence, especially with the self-social cognition module, highlighting the framework’s ability to capture the psychological impact of social exclusion.
The PSYA framework represents a significant step forward in simulating human-like behaviors, offering a richer and more accurate emotional and cognitive modeling approach for generative agents. This work provides a promising alternative to human participants in psychological experiments and has potential applications in simulating NPCs in games, virtual education, and social policy simulations. For more details, you can refer to the full research paper here.
Despite its advancements, the PSYA framework has limitations, including significant computational resource requirements for large-scale simulations and a still-developing capacity to capture the full spectrum of human emotions like ambivalence or regret. Future work will focus on optimizing performance and expanding the emotional complexity of the agents.
