Avalilação PREreview de Generative Agents: Interactive Simulacra of Human Behavior

This paper introduces generative agents, a novel class of interactive software agents designed to simulate believable human behavior using large language models augmented with memory, reflection, and planning mechanisms. The work is well motivated by the need for realistic human proxies in interactive systems such as simulations, social environments, and prototyping tools. Rather than focusing solely on task completion, the paper emphasizes long-term coherence, social interaction, and emergent behavior.

A major contribution of the paper is its agent architecture, which extends an LLM with a natural-language memory store that records experiences, synthesizes higher-level reflections, and dynamically retrieves relevant context to guide future actions. The interactive sandbox environment, inspired by The Sims, provides a compelling demonstration of how these components interact to produce believable individual and collective behaviors. The qualitative examples, such as autonomous social coordination around a planned event, effectively illustrate the system’s capabilities. The ablation study further strengthens the contribution by showing that observation, planning, and reflection each play a critical role.

However, the evaluation is largely qualitative and subjective, with limited quantitative measures of believability or robustness. Scalability beyond small populations and long-term stability under more complex environments are also not fully explored.

Overall, this paper represents a foundational contribution to agent-based AI research, offering architectural patterns that are likely to influence future work on interactive, memory-driven AI systems across domains.

Competing interests

The author declares that they have no competing interests.

Use of Artificial Intelligence (AI)

The author declares that they did not use generative AI to come up with new ideas for their review.