In a simulated town called Smallville, 25 AI characters go to work, remember their neighbors and make plans together. The project, described in the paper "Generative Agents: Interactive Simulacra of Human Behavior," uses large-scale language models to explore how believable social behavior might emerge from individual agents’ memories and decisions.
A town built from character details
The researchers from Google and Stanford University created a sandbox environment inspired by "The Sims." Each resident received a 1,000-character description covering their profession, personality and relationships with other residents. Those details give the agents a starting point for deciding what matters to them and how they might respond to people they know.
John Lin, for example, is described as a pharmacy shopkeeper who wants to make getting medication easier for customers. His background also includes his wife Mei Lin, a college professor, their son Eddy Lin, who studies music theory, and his connections to neighbors and colleagues. The descriptions give the characters social context before they begin interacting in the town.
Memory turns observations into plans
The agents’ behavior depends on a repeating process: they observe their surroundings, make plans and reflect on what has happened. They can also reflect on their own plans and earlier reflections. In the paper’s example, Klaus Müller comes to see himself as highly dedicated to research after accumulating and considering several observations.
To support this process, the system stores memories and retrieves them when useful. Memories are scored for recency, importance and relevance. A recent memory may be more useful than an older one; an event such as a breakup may matter more than breakfast; and a memory connected to the current situation may help guide what an agent does next.
ChatGPT handles these judgments using GPT-3.5 Turbo. To estimate importance, for example, it is asked to rate a memory’s poignancy on a scale of 1 to 10. The prompt contrasts mundane activities such as brushing teeth with events such as a breakup or college acceptance. This gives the simulation a way to distinguish routine details from experiences likely to shape future behavior.
Small interactions can add up
The researchers report that the agents showed what they call "believable individual and emergent social behaviors." One example began when a resident initiated a Valentine’s Day party. Within two virtual days, other characters invited themselves, arranged to meet and arrived at the agreed time and place. Some stayed away, reflecting the fact that not every character chose to attend.
The party illustrates how a social event can arise from agents acting on their own memories and plans. No single character needs to direct everyone else: an invitation can prompt further decisions, and relationships help make those decisions intelligible within the town. The result is a simulation of connected behavior rather than a set of isolated responses.
Potential uses come with social risks
The researchers see possible uses beyond a game-like demonstration. Generative agents could populate forums or virtual reality worlds for social prototyping. Combined with multimodal AI models, they might also contribute to social robots in the physical world. The team expects the overall architecture to continue working as language models develop, while saying that a model such as GPT-4 could improve the expressiveness and performance of the prompts beneath it.
More convincing agents could also blur the line between simulated interaction and a human relationship. The paper warns that people might form parasocial relationships with generative agents, anthropomorphize them or attribute emotions to them, even when they know the agents are computational entities. Developers should ensure that agents behave appropriately for their context, including avoiding responses such as reciprocating declarations of love.
The study’s evaluation period was short, and the researchers say longer experiments are needed to reveal the approach’s capabilities and limitations. Smallville offers an early look at how memory, reflection and social context can support AI characters that act together. How well those behaviors hold up over longer periods remains an open question.